Leveraging Machine Learning for Accelerated Electrode–Electrolyte Interface Design in Rechargeable Li‐Based Batteries

ABSTRACT

Due to their high specific energy, lithium–metal batteries (LMBs) are widely regarded as the promising next‐generation energy storage devices. Nevertheless, their practical applications are plagued by the challenges of irregular deposition and dissolution, coupled with the high chemical reactivity of lithium electrodes. Extensive research has focused on the stabilization of electrode–electrolyte interfaces as the key strategy to achieve improved battery performance. However, the exploration process via traditional “trial‐and‐error” methodologies is impeded by the long period and high cost of the tedious experiments. Machine learning (ML) technologies have become a mainstream force, redefining the revolutionary paradigm, enabling intelligently capturing the complex structure–performance relationships across vast compositional and structural spaces. Herein, ML applications in the discovery of electrolytes, electrodes, and interface engineering are reviewed, with the emphasis on ML‐driven investigation workflow covering data collection, feature engineering, model selection and ML‐assisted simulations. Moreover, task‐oriented ML technologies for expediting materials screening, informative descriptors extraction, mechanistic elucidation, and reverse design of novel electrodes and electrolytes are highlighted. Finally, future trajectories centered on multiscale materials simulation, multimodal modeling, and intelligent platform establishment to overcome persistent challenges are outlined, aiming at catalyzing the rational design of highly stable lithium electrode–electrolyte interface for long‐lasting rechargeable LMBs.

Rechargeable lithium–metal batteries (LMBs) with high specific energy have emerged as the promising next‐generation energy storage system. A closed‐loop machine learning (ML)‐driven framework covering data collection, feature engineering, model selection, and ML‐assisted simulations is highlighted to promote the design of electrode–electrolyte interface in LMBs. Future directions on comprehensive data library construction, multiscale simulation, multimodal modeling, and autonomous platform establishment are outlined.

1. Introduction

Since 1991, with the commercialization of lithium–ion batteries (LIBs), modern society has been transformed into a wireless, electrified, and low‐carbon system through a paradigm transition in communication and transportation infrastructure [1, 2]. Nevertheless, following over three decades of intensive development, conventional LIBs based on graphite anodes are reaching the intrinsic energy density limit of ∼300 Wh kg⁻¹ [3, 4], underscoring the urgent need to explore solutions to keep pace with the demand for increased energy density in long‐range transportation and grid energy storage applications. Impressively, Li–metal batteries (LMBs) demonstrate great promise owing to their exceptional theoretical specific capacity (3860 mAh g⁻¹) and the most negative electrochemical potential among any known anode materials (−3.04 V vs. the standard hydrogen electrode (SHE)) [5, 6]. Notably, recent advances in 11‐Ah‐level LMBs based on lithium anodes and engineered carbonate‐based gel‐solid‐state electrolytes deliver a record‐breaking cell‐level energy density of 604.2 Wh kg⁻¹ [7]. These breakthroughs position LMBs as a potential pathway to address the growing energy storage demands for broad future applications.

Despite their exceptional theoretical potential, developing Li metal anodes toward practical applications in batteries is fundamentally constrained by two interrelated challenges: formation of Li dendrites and the continuous parasitic reactivity at the anode–electrolyte interface [8, 9, 10]. Firstly, the inhomogeneous distribution of the electric field leads to the preferential reduction of Li⁺ ions at the tip sites of the electrode surface, giving rise to an irregular, even dendritic morphology. More seriously, uncontrolled dendritic growth can penetrate the separator, triggering battery short circuits [11, 12]. Meanwhile, during discharge, the nonuniform dissolution of lithium generates electrochemically isolated Li domains and microporous voids, irreversibly disconnecting conductive pathways and permanently trapping active lithium [7, 13, 14]. As a result, Li dendrite formation leads to severe safety hazards and significant capacity loss [15]. Besides, the high reactivity of Li electrodes is manifested in repeated interactions with electrolyte constituents. The extremely negative reduction potential of Li electrodes causes their thermodynamic instability in conventional organic electrolytes with relatively high lowest unoccupied molecular orbitals (LUMOs), which induces spontaneous reduction of solvent and salt species, forming a mechanically fragile and chemically heterogeneous solid electrolyte interface (SEI) [16]. Repeated volume fluctuations during cycling induce mechanical fracture of the SEI, exposing fresh lithium surfaces that trigger persistent parasitic reactions, leading to continuous electrolyte consumption [17], gas evolution, and progressive active lithium loss [5, 18]. This self‐accelerating cascade eventually lowers battery coulombic efficiency (CE) and reduces the specific energy [19].

To overcome these limitations, significant efforts have focused on manipulating the electrochemistry of the electrode–electrolyte interface (EEI) where both the deposition and dissolution of Li, and parasitic reactions occur. To address the instability issues in LMBs, a wide range of strategies have been explored, which can be broadly categorized into the electrode engineering (e.g., artificial SEI layer construction [20, 21]) and electrolyte design (e.g., electrolyte additives [22, 23], solid‐state electrolytes [24, 25, 26]). However, two fundamental knowledge gaps persist in the EEI design: (1) lack of a quantitative understanding of the relationship between interface characteristics (e.g., functional group of the SEI layer, crystal structure of the solid electrolyte) and interfacial stability [27, 28, 29]; (2) the unresolved in‐depth mechanism how atomic‐level interface features influence interfacial reaction kinetics and macroscopic battery performance [30]. Most studies rely on empirical, trial‐and‐error screening of parameter space, requiring protracted experimental research timelines and high cost, greatly hindering the translation of academic discoveries to industrial applications [31]. Moreover, investigation is often confined to a narrowly defined parameter space, inadvertently confining research in local performance optima and impeding the robust establishment of structure–performance relationships [7]. As the combinatorial space of chemical compositions, lattice symmetries, and surface terminations expands, identifying optimal regions in this high‐dimensional design landscape becomes intractable via conventional heuristics. Although established materials databases (e.g., the Inorganic Crystal Structure Database (ICSD), Materials Project (MP)) provide fundamental structural descriptors, the increasing novel components, such as metastable phases, multi‐interface layers and dynamically evolving SEI species, may deviate from well‐characterized structures, rendering exhaustive experimental research. Computational methods spanning density functional theory (DFT) and molecular dynamics (MD) simulations [32] efficiently accelerate the design and optimization of electrolytes and SEI layers by offering insights into electronic structures and Li⁺ migration barriers, as well as the atomic‐scale solvation dynamics and SEI nucleation pathways. However, the computational expense is prohibitive to simulate the long‐timescale interfacial dynamic processes involving electrolyte decomposition cascades, SEI formation, Li deposition and the solid–solid interfaces.

Machine learning (ML), fueled by advances in artificial intelligence (AI), constructs a new paradigm for materials research [33, 34, 35]. By training on the large‐scale datasets, ML enables the swift pinpointing of candidates with unprecedented efficiency. Crucially, ML techniques elucidate the correlation between compositions and structures with corresponding performances, and link computational predictions with experimental verification. This capability facilitates a deep mechanistic understanding of physicochemical principles underlying the EEI. Moreover, the establishment of the structure–performance relationship can make more accurate and knowledgeable decisions by the paradigm‐shifting integration of ML into electrode and interface layer design [36]. During this process, ML algorithms are capable of extracting powerful descriptors to forecast and promote the creation of novel solutions tailored to target performance. Furthermore, ML‐based interatomic potentials over quantum mechanical (QM) energy and force calculations have also proven highly successful, achieving both high accuracy and computational efficiency for performing structure optimization and evaluating thermodynamic stability, hence for advanced combinations design of elements and structures. Despite these advances, the structural optimization of electrodes and electrolytes for interface stabilization still relies heavily on trial‐and‐error approaches. ML‐driven strategies for interface stabilization in LMBs are still in the nascent stage. Thus, this underscores the need for a robust, data‐driven framework that can directly link meso‐scale structural attributes with macroscopic performance outcomes, enabling rational inverse design of advanced electrodes and electrolytes [37]. At the same time, several reviews have presented the achievements of ML‐driven battery research, centering on the specific stage such as high‐throughput (HT) technologies for data collection [38], algorithmic selection [39], ML‐based simulations [40], or unique applications such as AI‐aided electrode synthesis and smart manufacturing [41], solid‐state electrolyte property prediction [42], ML‐driven screening case studies for electrodes in LIB [43], batteries state prediction [44]. However, an end‐to‐end workflow encompassing ML‐driven screening frameworks and ML‐assisted simulations for lithium EEI engineering remains unaddressed.

Herein, different from existing review papers, this paper constructs four collaborative modules: (1) realizing an end‐to‐end closed‐loop ML‐assisted intelligent design workflow; (2) differing from mainstream research that focuses on bulk materials, it targets extracting effective descriptors to establish the unique design principle about the lithium metal EEI; (3) achieving cross‐scale integration to connect atomic, mesoscopic and macroscopic properties to elucidate the mechanism and guide the reverse design of high‐performance electrodes and electrolytes; (4) envisioning generative artificial intelligence with autonomous laboratories to enable autonomous generation and iterative verification of candidate materials. Roles of ML integrated into identifying the optimal candidates, extracting efficient descriptors, and elucidating mechanisms, as well as the achievements in electrodes and electrolytes design are discussed. Guided by these perspectives, the outstanding challenges and forward‐looking research directions in ML‐assisted interface engineering are presented, primarily addressing advancements in multiscale materials simulation, multimodal modeling, and the engineering of intelligent platforms, with the aim of advancing LMBs and establishing an exploration paradigm that promotes the rational inverse design of highly stable metallic electrode interfaces.

2. Advances in ML Techniques

Relying heavily on iterative trial and error, conventional research methods in both experiment and computation typically entail extended development periods and substantial costs [45]. By virtue of its capability of handling high‐dimensional data, ML has demonstrated an important role in discovering and designing electrodes and electrolytes by predictively learning structure–property links. These models are subsequently harnessed within an optimization framework to guide the quest for superior interfacial chemistries or formation protocols. In practical use, ML techniques can be coupled with experimental techniques (e.g., X‐ray photoelectron spectroscopy, scanning electron microscopes and electrochemical performance testing) and theoretical calculations (e.g., density functional theory, ab initio molecular dynamics (AIMD) simulations) [46]. Accordingly, the applications of ML are categorized into two primary modes: establishing a screening pipeline based on structure–property relationships, and performing MD simulations based on machine learning force fields (MLFFs).

2.1. Establishing ML‐Based Framework for Structure–Performance Relationships

Expediting the discovery and design of electrodes and electrolytes is pivotal for propelling the development of LMBs. Linking material structure and composition to their intrinsic properties to establish robust structure–property relationships remains the mainstream method for accelerating materials discovery. Supported by this framework, a predictive pipeline is established to enable streamlined screening and refinement of candidate materials, which is generally based on the critical stages of data collection, feature engineering and model selection (Figure 1). High‐quality, diverse, and representative training data is key to model performance and generalization. Through careful data analysis and feature engineering, models can be guided to extract physically meaningful trends instead of spurious relationships. This framework further supports a predictive pipeline that efficiently screens and optimizes candidate materials. Finally, rational algorithm selection is highly demanded to tackle the complex and heterogeneous data.

FIGURE 1.

Workflow of ML–assisted interface engineering.

2.1.1. Data Collection

Within the ML workflow, the initial and critical step lies in data collection, which provides an input foundation for subsequent model training and performance optimization. Broadly speaking, data is classified into two basic types: target properties and structural/compositional features. Target properties represent the physical or chemical quantities that the model aims to predict. They reflect the key properties of materials or molecules, including electrolyte ionic conductivity and the mechanical stability of SEI. This type of data also includes energy and force corresponding to atomic configurations, Li⁺ migration energy barriers, battery cycle life, and critical temperature of superconductors. The feature data is used to describe the characteristics of the researched materials, such as compositional formulations, structural information of atomic coordinates, lattice parameters, stoichiometric ratios at the atomic scale, as well as the process descriptors [47], providing the ML model with a basis for learning the correlation between materials and properties. In general, data can be derived from the established database, experimental, or computational dataset (Figure 2). Extensive materials databases that integrate millions of structural records with both experimental and DFT‐calculated properties (e.g., crystal structures, band gap, formation energy) have been established, such as ICSD, MP, Novel Materials Discovery, Automatic Flow for Materials Discovery, and Open Quantum Materials Database (Tables 1 and 2). These structured or semi‐structured data can be easily structured for downstream processing [48]. Aside from the aforementioned general databases, various platforms specifically built for battery electrolytes have emerged and achieved significant progress. Simply put, such platforms focus on three major categories: molecular systems, solution systems, and solid‐state electrolytes. Most of these platforms are constructed by integrating theoretical calculations. For molecular‐level design, a typical example is the Electrolyte Genome project. It focuses on a vast number of battery electrolyte molecules and contains data on 4830 electrolyte‐related molecules (including ionization potentials and electron affinities). It has performed large quantum‐mechanical calculations on molecular properties via HT quantum computation, and integrated corresponding automatic fault repair programs and reliable structure matching tools, which greatly improves the efficiency and accuracy of candidate molecule screening [49]. Simultaneously, the research teams led by Chen and Zhang from Tsinghua University have jointly developed the Uni‐Electrolyte platform, an AI design system for secondary battery electrolyte molecules, together with the Beijing AI for Science Platform. This platform consists of three collaborative modules and integrates multiple functions including novel molecule design (EMolCurator), retrosynthetic pathway planning (EMolForger) and interface evolution analysis (EMolNetKnittor). It can comprehensively evaluate various key properties of electrolytes, so as to realize the intelligent discovery of electrolyte molecules, synthesis assessment and interfacial stability prediction [50]. Currently, research on electrolyte‐related data no longer merely focuses on single molecules, but pays more attention to the overall properties of solutions. Gregory et al. [51] established the IonSolvR ion solvation database. A large number of molecular dynamics (MD) simulations were carried out using the DFTB method, generating more than 3000 simulation trajectories. The database covered combinations of 55 types of ions and 28 different solvents. Each trajectory contained microscopic structural information such as coordination number, ion–solvent interaction energy, and radial distribution function. With high precision, it can serve as a reliable reference for exploring the laws of ion dissolution and migration at the solution level. Nakamura et al. [52] built another type of open‐source battery electrolyte database named OEDB. Relying on fully automated HT simulation methods, this database collected nearly 5600 electrolyte formulas, including 3 cations, 9 anions, 26 solvents and solutions with various concentration ranges. It summarized key bulk properties of solutions such as viscosity, ionic conductivity and density. Supporting online visual viewing via an interactive web interface and intelligent retrieval, the database facilitated electrolyte formula screening and multi‐objective optimization. Furthermore, remarkable progress has also been made in the construction of databases for solid electrolytes. For instance, the OBELiX database system compiles room‐temperature experimental ionic conductivity and crystal structure information of approximately 600 solid electrolyte materials, among which complete Crystallographic Information File (CIF) documents are available for around 320 materials. It provides high‐quality labeled data for the research on the structure‐performance relationship of solid electrolytes [53]. The literature‐derived migration barrier dataset extracts 621 independent ion migration barrier data points calculated by the DFT‐NEB method for 443 electrode and solid electrolyte materials from published literature. It also offers detailed information about the initial and final positions of ions as well as corresponding energy barriers, which can be directly applied to ML models requiring structural input such as GNNs [54]. Recent review articles have systematically sorted out the full set of application schemes of ML in the research and development of various electrolytes. They cover the whole‐chain technical paradigm ranging from additive screening to molecular property prediction, multi‐objective model construction, and closed‐loop autonomous experimental optimization for formulation design, breaking down the barriers between computational simulation and experimental research and development [55]. The introduction of the above‐mentioned specialized electrolyte databases, informatics platforms and new data types makes up for the deficiencies of general databases, enriching the research of ML‐assisted electrolyte design and bringing it closer to practical application scenarios. Nevertheless, the unlabeled data in these databases is difficult to harness. Experimental data is directly obtained through characterized structures and tested performances in the realistic environment, which can be performed by laboratory manual experiments and HT experiments. Encouragingly, HT experiments synthesized and screened candidates in a parallel manner, greatly facilitating the construction of databases correlating structural parameters with output metrics [56]. However, navigating the vast parameter space still faces the challenges of long research time and high expense. In comparison, calculations via DFT or AIMD simulations break through the limitations of experimental conditions, improve the scale of data, but also provide critical insights into the atomic‐scale structures (e.g., coordination environment, Li⁺ ions solvation structure, thermodynamic energy) [57, 58, 59, 60]. However, the calculation results of some complex systems may deviate from actual experiments, which lowers the predictive accuracy of ML models. To tackle the deficiencies, it is crucial to construct a multi‐dimensional database for electrodes and electrolytes screening that integrates knowledge domain database, experimental and simulation results. Additionally, it is worth noting that supplementary data, which focuses on the easily overlooked low‐performance data or failed experimental data also deserves much attention. It can effectively alleviate the bias of ML models, improve their generalization ability, make up for data value deviation, avoid overfitting of ML models to high‐performance samples, and thus expand the coverage of models to the real material space. A typical example is the research by Raccuglia et al. in which experimental data collected from failed hydrothermal synthesis reactions was integrated to train a ML model for predicting the crystallization of templated vanadium selenate crystals. It was found that the performance of this model was superior to that of traditional manual analysis, with an accuracy rate of 89% regarding the formation conditions of novel organic‐templated inorganic products [61]. In summary, the predictive performance of ML models is fundamentally governed by the volume and integrity of the training data. Larger volumes of high‐quality data directly translate to more accurate models and more precise predictions. Therefore, it is highly demanded to enhance database robustness characterized by experimental reproducibility, accurate labeling, and comprehensive feature coverage.

FIGURE 2.

Data sources for ML research.

TABLE 1.

Examples of experimental databases [62].

Database name	Core data content	Application scenarios
ICSD [63]	Over 190 000 inorganic crystal structure entries, including lattice parameters and atomic coordinates. Technical errors have been checked and eliminated, but human /measurement errors still exist	Crystal structure‐related research (such as crystal structure prediction, lattice stability analysis)
CSD [64]	Crystal structure data of organic /metal–organic compounds, including structural information such as bond lengths and coordination numbers	Molecular‐scale materials design (such as catalyst structure optimization, organic electrolyte molecule design)
Battery DB [65]	Manufacturer specification data of over 500 commercial secondary batteries, including electrode/electrolyte formulations and cycle life	Screening and performance prediction of commercial battery materials
HTEM Database [66]	Data of 57 597 thin‐film samples, covering oxides, nitrides, etc., including composition and structural phase information	Research on thin‐film battery materials (such as electrode thin films for solid‐state batteries)

TABLE 2.

Examples of computational databases.

Database name	Core data content	Application scenarios
MP [67]	Over 133 000 DFT‐relaxed crystal structures, including energy, band gap, elastic constants, etc., with supporting API for quick access	Screening of inorganic battery materials (cathodes, anodes, solid electrolytes)
AFLOWLIB [68]	Over 3 million inorganic entries and 7 × 108 calculated properties (based on high‐throughput DFT), including lattice energy and electronic conductivity	Prediction of candidate materials for battery electrodes /electrolytes
OQMD [69]	Data of over 560 000 computationally predicted materials, including formation energy and stability parameters, expanded based on ICSD experimental structures	Discovery of new inorganic materials (sulfide electrolytes)
NOMAD Repository [70]	Computational materials property data (energy, force, electronic structure), supporting the FAIR principles	Cross‐domain materials data reuse (extending from battery materials to catalytic materials)

Currently, the data sources for ML‐based battery research fall into three categories: public databases, HT experiments, and theoretical calculations, which have greatly increased data volume and acquisition efficiency. However, data from different sources suffer from prominent quality disparities, systematic deviations and inconsistent noise, severely degrading the performance of fusion training on multi‐source data. Experimental data are vulnerable to fluctuations in sample purity, equipment errors and testing conditions, accompanied by random noise and poor repeatability. Even when the same materials are tested with identical methods, the performance values reported in different publications vary considerably, leading to inherent data uncertainty [47, 61]. Although data obtained via DFT theoretical calculations feature standardized structures and large sample sizes, they are constrained by parameters such as functional selection, cutoff energy and model approximations, resulting in systematic deviations from actual experimental results. Direct joint training on the two types of data will markedly impair the generalization ability of models [71]. Meanwhile, existing public databases are plagued by inconsistent formats, strong heterogeneity and high commercial barriers, making cross‐database data alignment and joint training difficult [72] and further complicating the fusion of multi‐source data. To address the quality defects and systematic deviations of such multi‐source heterogeneous data, a set of mature strategies for data cleaning, standardization and uncertainty quantification has been developed. Data cleaning activities can be divided into six modules: feature cleaning, label cleaning, entity matching, outlier detection, missing value imputation and global data processing [73]. They can be combined with unsupervised algorithms such as Principal Component Analysis (PCA) and K‐means clustering to identify anomalies, and statistical methods including the Z‐score are adopted to quantify the degree of sample deviation, so as to reduce data noise and alleviate model overfitting. Data standardization takes both data itself and data management into account. StandardScaler can be utilized to standardize data prior to analysis to eliminate dimensional differences among different attributes [74]. Meanwhile, labeling specifications are unified in accordance with the FAIR (Findability, Accessibility, Interoperability and Reusability) principles to connect different databases and realize data sharing and integration [75]. For the problem of severe discrete fluctuations in experimental data, uncertainty quantification adopts statistical indicators such as the median, which are insensitive to extreme values, to accurately quantify inherent errors in data and improve the confidence level of modeling laws [76]. A reliable high‐quality integrated dataset is ultimately constructed. Based on the above preprocessing, we can further rely on active learning to iteratively screen highly uncertain samples for precise supplementary labeling. Combined with transfer learning pre‐training strategies, this approach narrows the data distribution gap between the source domain and target domain, effectively mitigating the misalignment between high‐fidelity and low‐fidelity data [77]. When facing data sparsity and insufficient samples, generative models such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) can generate high‐quality augmented data. Together with multi‐physics‐field simulations to expand samples with diverse configurations, they further enrich the training space.

The future development of data systems will focus more on the standardized alignment and multi‐fidelity integration of multi‐source data [78]. By establishing unified and open databases, promoting data sharing between industry and academia, and adopting physics‐constrained data augmentation technologies as well as domain‐optimized descriptors, researchers will gradually resolve core challenges including mismatches between experimental and theoretical data, high uncertainty and poor transferability. This will provide high‐quality data support for the ML screening of electrolyte and electrode materials.

2.1.2. Feature Engineering

Typically, researchers rely on manual design (also referred to as domain knowledge‐driven) or physical mechanism‐driven methods for feature construction. Extracting features from the experimental characterization data is the straightforward method. For example, extracting the contents of LiF and Li₂CO₃ from XPS data of the SEI film [79], and extracting the number of fluoro groups (‐F) and ether linkages (ROR') from electrolyte molecular structures [80]. In terms of the mechanism‐driven method, it constructs features based on the physical and chemical laws of interface reactions. A representative example is the use of the Arrhenius relationship [81] to convert temperature–interface impedance into interface activation energy, which is then correlated with SEI growth kinetics. Despite their common use, these methods have intrinsic limitations: manual design fails to distinguish materials with the same composition but different structures, introducing biases that may lead to suboptimal performance and create cognitive blind spots for new insights [72]; the physical mechanism approach requires formula derivation, which complicates efforts to quantify the coupled effects of multiple factors [82], and this process is time‐consuming and prone to errors [83]. Notably, automated feature extraction methods based on deep learning directly learn features from raw inputs, thereby avoiding the subjective biases of manual design and reducing reliance on domain knowledge [84, 85]. Meanwhile, it can capture information that is difficult to obtain via manual or physical methods, enabling the characterization of complex structures and nonlinear correlations.

The collected data encompasses high‐dimensional structures and properties, which is inherently prone to overfitting. To mitigate this risk, feature engineering is of high significance to eliminate redundant or minimally informative predictors and construct informative descriptors (Figure 3). Such strategies are highly important to strengthen the statistical robustness of learning algorithms in data‐scarce regimes, analyze the intrinsic connections between these original features [86], thereby improving the discovery efficiency and accuracy.

FIGURE 3.

The workflow of ML‐driven feature engineering.

In practical use, multiple feature transformation approaches have emerged. Among the various transformation strategies, physical mapping transformation is simple and straightforward, which is based on known physical and chemical laws to convert key attributes that cannot be directly measured into computable numerical features based on raw data. For example, Liu et al. [82] used coulomb matrix, a physical tool, to convert nuclear charges and interatomic distances into matrix features, where the diagonal elements were used to represent atomic energy, and the off‐diagonal elements were used to represent electrostatic interactions between atoms, thereby fully encoding the 3D structure of electrolyte crystals. Nevertheless, it is highly dependent on known physical laws, which is limited in analyzing unknown side reactions or multi‐field coupling phenomena without well‐established physical formulas [84]. The symbolic regression transformation based on algorithms such as Genetic Programming for Symbolic Regression (GPSR) can automatically find patterns from scattered original features in interface design and generate formula‐based interface features, thereby realizing the quantitative expression of multi‐factor nonlinear coupling [87]. As demonstrated, composite features from original features such as the tolerance factor and octahedral factor of perovskites were extracted via symbolic regression [88]. However, a critical point is that its performance largely hinges on access to a significant amount of data. If the dataset is too small, it is easy to cause model overfitting. Graphical transformation can be used to solve scenarios that are difficult to handle by the first two methods. Its principle is to convert structured numerical data (such as Coulomb matrices and impedance spectra) in interface design into 2D image formats, and then use models such as Convolutional Neural Networks (CNNs) to extract relevant features related to space and morphology [89]. For example, Liu et al. first converted a 202 × 202 symmetric Coulomb matrix into a Coulomb Matrix Graph (CMG) image, mapping the values of matrix elements to pixel grayscale values, and converting the spatial distribution of electrostatic interactions between atoms into image textures. Subsequently, the CNN model successfully extracted local structural features [83]. Despite the great promise, graphical transformation significantly relies on image analysis models and sufficient image data for model training, which is limited to analyze the precise numerical information. Collectively, there is a trade‐off between the accuracy requirements and the existing resource conditions. In practical applications, the selection of feature transformation methods should be guided by the underlying physical nature of the system and the availability of domain‐specific data. For scenarios supported by clear physical laws, such as the prediction of Li⁺ diffusion coefficient in dilute electrolytes of traditional liquid electrolytes where Mean Squared Displacement provides a direct mechanistic link to ion mobility [90], and the unit cell structure optimization of solid electrolytes with regular structures (Coulomb matrix can be used) [91], physical mapping transformation is preferred due to their interpretability and physical consistency. In contrast, for systems lacking explicit physical models but characterized by abundant high‐dimensional data, such as the prediction of new electrolytes and electrodes, ion aggregation behavior, symbolic regression emerges as a suitable approach, enabling the automated discovery of interpretable mathematical expressions from data. For scenarios involving spatially resolved, morphologically complex interfacial data, such as microscopic structure images of EEIs, and scenarios requiring high‐precision detail recognition such as grain boundary defect detection of solid electrolytes, graphical transformation demonstrates superior performance.

Following feature transformation, features with negligible correlation to the target are often pruned, given their limited predictive value. One method to identify such features is correlation analysis, which quantifies their relationship with the target. The filter method is a rapid screening approach based on statistical indicators. The commonly used indicators include Pearson Correlation Coefficient (PCC) and Maximal Information Coefficient (MIC) [92]. For example, ranging from −1 to 1, PCC is a widely used metric: values near the extremes (±1) mark strong linear dependence, in contrast to values near 0, indicating a lack thereof [93]. If a feature's Pearson correlation substantially falls below a preset threshold, removing it can lower dimensionality and mitigate noise in subsequent modeling. Additionally, researchers have proposed the regularization method, a type of feature screening constrained by loss functions [72]. It refers to adding a penalty term to the model's loss function and achieves the integration of feature selection and model training by penalizing the absolute values or the sum of squares of feature coefficients. This method can resolve the issues of high‐dimensional redundancy and multicollinearity; however, it relies heavily on data preprocessing and results in models with poor interpretability. On this basis, to facilitate better subsequent model selection, the dimensionality reduction method has been put forward. As a low‐dimensional mapping technique for high‐dimensional features, it can map high‐dimensional features to a low‐dimensional space through linear or nonlinear transformations, retaining core information. To achieve more effective feature selection, it is highly demanded to perform feature filter based on the applications scenarios. In general, the filter method can be adopted for the initial feature screening of liquid electrolytes and solid electrolytes [94, 95], especially for the preliminary screening of molecular structure features and element attribute features. The regularization method can reach the feature screening of new‐type electrolytes or the accurate screening of element attribute features and molecular structure features [96, 97]; while the dimensionality reduction method is more suitable for feature processing of solid electrolytes and multi‐modal data, along with the dimensionality reduction of high‐dimensional simulation data [98, 99]. Additionally, to achieve more accurate feature selection, the aforementioned methods can also be used in combination. Specifically, low‐value features can first be quickly eliminated via the filter method, and then, based on the sample size and feature correlation, either the regularization method or the dimensionality reduction method can be selected for further optimized screening.

Once irrelevant features are filtered out, it is equally vital to perform optimization and verification to achieve accurate feature adaptation and performance assurance. At this stage, correlation analysis can once more be applied to evaluate how strongly features correlate with the target performance. To illustrate, Yan et al. [80] used a Random Forest (RF) model as the core verification tool, training the RF model with the screened key features, and finally proved that the screened features exhibited high prediction accuracy, achieving a high coefficient of determination (R ² = 0.7038) and a low mean squared error (MSE = 0.0074). Besides, attention should be paid to develop the corresponding evaluation method to enhance model accuracy. As reported by Shin et al. [88], descriptors based on MIC were found to exhibit better correlation performance than those based on PCC, with an accuracy even exceeding 90%. In this case, MIC can be prioritized for nonlinear correlations of interface performance.

In the design of EEIs, feature engineering serves as the core support for the practical application of ML. Its essence lies in being guided by interfacial physical mechanisms and integrating data‐driven approaches with domain knowledge. This type of feature engineering paradigm can be further combined with multi‐scale interfacial data ranging from molecular simulations to real‐world characterization, providing more efficient ML tools for the interface optimization of LMBs.

2.1.3. Model Selection

Model selection refers to the process of selecting a suitable model for the current problem from different algorithm categories and specific models in ML by combining task objectives, data characteristics and resource constraints. The well‐constructed ML models are critical to extract informative descriptors, or forecast structure–performance connections [100].

The selection of ML models places great emphasis on the need for alignment between model capabilities and problem requirements. Different models have clear boundaries of application, and incorrect selection can lead to performance failure. The ML models can be generally categorized into supervised, semi‐supervised and unsupervised based on the amount and characteristics of the dataset (Figure 4). As the most powerful and mature approaches, supervised learning (SL) is widely used in ML investigation. SL learns from labeled training data (input–output examples) in order to derive a function that can generalize and accurately predict outputs for unseen inputs [41]. SL algorithms encompass regression models and classification models, which differ in the type of predicted values. Typically, regression models are used to predict continuous numerical values, including linear regression, Gaussian Process Regression (GPR), Gradient Boosting Regression (GBR)/XGBoost/LightGBM [101, 102] and Random Forest Regression [103]. In contrast, classification models are favorable to predict discrete categories, including logistic regression [83], Support Vector Machines [104], Random Forest Classifiers [105], GNNs, and CatBoost [106]. If the available dataset consists of only input values, unsupervised learning (UL) can be used in an attempt to identify previously unrecognized patterns in large datasets, patterns or clustering in the data. UL can process unlabeled samples and mainly consists of clustering and dimensionality reduction algorithms. Examples include the K‐Nearest Neighbors algorithm, which is simple, efficient, and suitable for data grouping (e.g., material subset division), and Principal Component Analysis, which minimizes information loss and improves data interpretability (e.g., for preprocessing high‐dimensional spectral data) [107]. Semi‐supervised learning (SSL) is superior in scenarios with abundant input data yet limited matching output labels [108].

FIGURE 4.

ML algorithm categories.

The core basis of model selection lies in aligning with task requirements, data characteristics, and constraints of practical applications. In model selection for applications, defining, or properly framing, the predictive task is the pivotal step, such as identifying the promising candidates, elucidating the reaction mechanism, or performance prediction. Subsequently, a rigorous evaluation of the available data is imperative. This includes evaluating the dimensionality and type of features: if the features are high‐dimensional discrete ones containing category information or special types such as atomic and molecular structures, whether the data volume is large (with tens of thousands of samples) or small (with hundreds of samples), whether there is labeled information, and whether there are missing or messy data. A typical selection strategy can be referenced as follow. Generally, if there is labeled data and the goal is to predict continuous values, regression models in SL often prove adequate, directly utilizing the labeled data without extensive preprocessing. By contrast, when there is unlabeled data, clustering models in UL are often prioritized [109], because of their effectiveness in quickly identifying groups of similar data. Interpretable linear models such as Logistic Regression and ensemble methods like RF are efficient to tackle the high‐dimensional discrete feature spaces, which balance performance with feature importance transparency. Additionally, Light GBM and deep neural networks that are employed to leverage computational efficiency and representational power are efficient for large‐scale datasets (tens of thousands of samples). Conversely, for small‐sample regimes (hundreds of samples), GPR and multi‐task learning frameworks are adopted to mitigate overfitting and exploit shared latent patterns across related tasks. In a more specific scenario, when modeling atomic or molecular structures represented as graphs, GNNs are favored for their ability to capture nonlocal structural dependencies and quantify predictive uncertainty. Under strict time or resource constraints, lightweight models such as Logistic Regression and K‐Nearest Neighbors are preferred for rapid inference.

Following the model construction, cross‐validation of the established ML models should be carried out to test the efficiency, generalization and applicability in practice. For example, 10‐fold cross‐validation was repeated 3 times to reduce random errors and ensure stability across different data partitions for the large‐sample data [102]. For small‐sample data, the leave‐one‐out cross‐validation (LOOCV) method is employed. Only one sample is used as the test set, while all others collectively form the training set, with all samples tested sequentially to avoid wasting valuable data, as demonstrated in the ML‐driven prediction of solid‐state electrolytes (SSEs) [101, 110]. Additionally, external validation using independent data outside the training set deserves much attention, such as data from new experiments or DFT simulations. A representative example is that Crystal Graph Convolutional Neural Networks (CGCNN) determines the key attributes of MOF materials with accuracy comparable to DFT, illustrating the reliability of the chosen model [108]. Besides, predictive capability of the model can be evaluated by the correlation analysis based on Mean Absolute Error (MAE), Mean Absolute Percentage Error, Matthews Correlation Coefficient. The model's core hyperparameters are then optimized by methods like Bayesian optimization to improve the performance of the model. It was demonstrated that the accuracy of LOOCV can reach 0.903 after optimizing LGBM with Bayesian optimization [101]. Crucially, all selected models should be subjected to physical plausibility screening to enforce adherence to fundamental electrochemical principles [111]. As demonstrated, SHAP analysis revealed that relative density and the electronegativity of Li‐site doped elements are key factors affecting ionic conductivity. This is consistent with the common sense that denser materials have lower resistance.

The above discussion centers primarily on post‐hoc physical screening methods as a means to validate the prediction accuracy of traditional data‐driven ML models in quantitative structure–activity relationships. Even so, for complex systems with strong nonlinearity and multi‐field coupling such as EEI, prioritizing predictive accuracy alone leads to a lack of intuitive interpretability between inputs and outputs, namely the “black box” dilemma. Consequently, the design guidelines derived through such paradigms also suffer from poor transferability [81]. For this reason, Physics‐Informed Machine Learning (PIML) and Explainable Artificial Intelligence (XAI) have emerged as key approaches to mitigate existing bottlenecks [48, 112]. The core idea of PIML is to embed fundamental physical laws (such as mass conservation and reaction kinetics) as prior information into model training, instead of relying solely on pure data fitting. For instance, Physics‐Informed Neural Networks (PINNs) directly incorporate the residuals of partial differential equations (including Fick's law for describing Li⁺ diffusion and reaction‐diffusion equations for depicting SEI growth) into the loss function. This enables the model to maintain physical consistency even in data‐sparse regions and realize physically constrained prediction of interface evolution [113]. Meanwhile, XAI is committed to breaking the “black box” barrier, allowing researchers to understand the causal relationships between features and model outputs. Among various XAI methods, SHAP, based on the game theory‐derived Shapley value, uniquely determines each feature's contribution when satisfying local accuracy, missingness, and consistency. It has thus become the most widely applied interpretation framework in current battery material research [114]. Furthermore, if PIML can be synergistically integrated with XAI, this combination will not only improve predictive precision but also extract descriptors with clear physical meanings and verify the rationality of model decisions, laying a foundation for establishing transferable interface design criteria.

Despite these ongoing advances, core challenges persist in applying ML to materials science, such as heterogeneous data quality, the inherent difficulty of achieving full interpretability in complex systems and poor generalization performance on small datasets [115]. In the future, with the optimization of feature engineering techniques, cross‐domain data integration, and iteration of advanced algorithms, ML is expected to break through the limitations of traditional research models in more scientific scenarios and continuously unleash its innovative potential [116].

2.2. Performing Simulations

To gain an in‐depth understanding of the mechanism for electrochemical reactions at the EEI, machine learning potentials (MLPs) have been extensively adopted, which are highly specialized in the field of atomic simulation, representing a state‐of‐the‐art approach to exploring the thermodynamic mechanisms in materials. To be specific, MLP technologies employ ML algorithms to learn from quantum chemistry calculation data to analyze the underlying potential energy surface (PES) that governs materials behaviors, thereby mapping multi‐dimensional structures to corresponding energies and forces [113].

The operation of MLPs begins with the mathematical abstraction of the physical world. Unlike predefined physical functions such as the Lennard‐Jones potential, MLPs utilize highly flexible mathematical functions (e.g., neural networks) to construct a correlation linking atomic structures to the total system energy [117]. They decompose the total energy E into contributions from individual atoms to enhance interpretability and training efficiency, with their core design paradigm being the decomposition of the entire system's potential energy into the aggregate of local energy terms from each atom.

$$E={\sum }_{i=1}^N{E}_i$$ (1)

$$E_i=M\left(\left\{r_{ij}\right\},j\in K\left(i\right)\right)$$ (2)

where the energy E_i of each atom is determined solely by its local chemical environment, that is, the relative coordinates r_ij = r_i − r_j of its neighboring atoms K(i). M denotes a ML model, such as GNNs and large language models (LLMs). This locality assumption serves as the foundation for the efficient scalability of the model, while atomic forces are derived by calculating the negative gradient of the potential energy through automatic differentiation techniques

$$F_i=-{\nabla }_{r_i}E$$ (3)

thereby strictly ensuring energy conservation. Here, the MLP acts as the core for the parametric representation of the potential energy function, while the MLFF constitutes a complete computational framework that integrates this potential energy model to drive simulations for mapping correlations between system energy and atomic positions or charges.

Constructing a robust and reliable MLP model is a systematic and iterative engineering process, with its core lying in establishing a high‐precision and high‐efficiency atomic configurations. As illustrated in Figure 5, the entire workflow commences with data generation, where an initial set of physically representative atomic configurations for the target system must be designed. These configurations are typically derived from AIMD simulations, enhanced sampling techniques, or pre‐sampling based on existing empirical force fields [118]. Each configuration is labeled via QM calculations (e.g., DFT) to obtain reference values of total energy, atomic forces, and virial stress, which then serve as training benchmarks. Subsequently, it is demanded to convert the three‐dimensional coordinates of atoms into mathematical features that are both processable by ML models and strictly preserve the translational, rotational, and permutational invariances required by physical systems. For example, GNNs provide a natural representation by treating atoms as graph nodes (with node states containing information such as element identity) and atomic connections as edges (carrying geometric information such as distances), then iteratively propagating information across the graph via learnable message‐passing functions to ultimately encode the local chemical environment of each atom. Going forward, equivariant neural networks can directly construct tensor features that follow deterministic transformation rules under rotational operations, thereby enabling a more elegant and efficient handling of vector outputs such as atomic forces. Moreover, in recent years, researchers have even attempted to adapt transformer architectures for atomic systems,

$$a_{ij}=\frac{\exp \left(\frac{q_i^T{k}_j}{\sqrt{d}}\right)}{{\sum }_{j\in K\left(i\right)}\exp \left(\frac{q_i^T{k}_j}{\sqrt{d}}\right)},h_i^{\prime }={\sum }_{j\in K\left(i\right)}{a}_{ij}{v}_j$$ (4)

where q_i = W_Qh_i (queries), k_j = W_kh_j (keys), v_j = W_Vh_j (values) and d represents the key dimension. By treating atomic or structural features as tokens, mechanisms are employed to model complex interactions. Upon obtaining atomic environment feature representations that satisfy symmetry requirements, the workflow proceeds to the nonlinear fitting stage, where these features are fed into a deep neural network. Loss functions are designed with physics‐guided regularization, such as an energy‐force coupled loss

$$\mathcal{L}={\lambda }_E{‖E_{ML}-E_{QM}‖}^2+{\lambda }_F{\sum }_i{‖F_i^{ML}-F_i^{QM}‖}^2$$ (5)

where E_QM and F_i ^QM are QM‐calculated reference values, and λ_E and λ_F are weighting coefficients for the energy and force accuracy [113]. This function can simultaneously constrain the prediction errors of energy, atomic forces, and even stress, with hyperparameters used to balance the weights of each term, ensuring that the trained model not only achieves accurate total energy predictions but also yields highly reliable atomic forces (i.e., the negative gradient of energy). However, since the initial training set can hardly exhaust all physically relevant configuration spaces, it is necessary to expand the training set through structure‐based or property‐based selection techniques (or a combination of both) and repeat the process until satisfactory model performance is achieved. Once an initial model is obtained, it can be readily used to generate new structures, often accelerating the process of structural exploration. Finally, when the developed model passes all application‐related tests, it can be applied to provide insights into challenging scientific problems. In addition, a crucial technical issue is addressing long‐range interactions that extend beyond the local cutoff radius. Traditional methods based on local environment descriptors ignore nonlocal and long‐range interactions between atoms, leading to significant errors especially in systems with strong charge transfer or van der Waals interactions [119]. To tackle this problem, the mainstream solution adopts a physics‐informed hybrid modeling strategy that decomposes the overall energy into short‐range and long‐range components. Short‐range strong interactions such as chemical bonding are flexibly fitted by neural networks, while long‐range components are calculated using analytical methods based on physical laws. For instance, a common practice is to use a small neural network submodule to predict the partial charges of each atom and then compute the global electrostatic interaction energy via the classical Ewald summation technique. More sophisticated schemes introduce learnable atomic electronegativity parameters and self‐consistently determine atomic charge distributions by solving global charge balance equations, thereby enabling more accurate simulations of charge transfer effects. Cutting‐edge research has also attempted to use ML to directly predict atomic dipole moments and polarizabilities, or to characterize the polarization response of electron clouds using maximally localized Wannier function centers (MLWCs), achieving a more refined description of long‐range interactions [120].

FIGURE 5.

Integrated ML workflow for property‐aware modeling of materials.

Through the above complete and rigorous technical workflow encompassing data design, symmetry‐preserving feature engineering, iterative training integrated with active learning (AL), and physics‐informed hybrid modeling, a successfully trained MLP can realize its revolutionary value. It performs large‐scale molecular dynamics simulations with accuracy comparable to ab initio calculations, while achieving computational efficiency millions of times higher than the latter [113]. This makes it possible to investigate scientific problems that were previously inaccessible to traditional methods on spatiotemporal scales of millions of atoms and nanoseconds or even microseconds, such as the long‐range migration of ions in SSEs and complex reaction networks on heterogeneous catalytic surfaces. Furthermore, HT simulations based on MLPs have themselves become incubators for new scientific concepts, with the new conceptual framework termed density of atomistic states (DOAS) providing a novel analytical tool for quantitatively understanding the energy landscapes of complex disordered systems [121]. Data‐driven and automated construction methods have significantly reduced the number of high‐cost reference calculations required for model development, enabling the adoption of higher‐level electronic structure techniques as training benchmarks and thus continuously pushing the upper limit of model accuracy.

By contrast, their workflow still faces great challenges. In terms of accuracy boundaries, integrating data from more advanced quantum chemical methods (e.g., coupled cluster theory) at an affordable cost requires transfer learning (TL) techniques such as delta learning. In terms of generalizability, most MLPs remain specialists tailored to specific systems; developing “lifelong learning” potentials that can generalize across elements and compounds while enabling continuous learning without forgetting represents a key research direction [122]. Meanwhile, the accurate description of strongly correlated systems, excited states, and extreme chemical bonding conditions, as well as ultimately closing the remaining gap between simulations and experimental observations, remain unsolved problems.

Beyond the above common challenges, there are additional limitations when applying traditional local ML methods to strongly charged electrochemical interface systems such as lithium metal anodes. They fail to introduce realistic electrode potential boundary conditions, and thus can hardly accurately characterize actual interfacial behaviors including ion accumulation, charge redistribution and uneven deposition of lithium metal. To address this limitation, two major technical approaches have been developed in academia in recent years, namely the charge equilibration method and the charge transfer method, which break the simulation boundaries of traditional MLFFs from different physical perspectives. The core idea of the charge equilibration method is to dynamically adjust the charges of atoms on the electrode surface, so that the system can reach electrochemical equilibrium under a given electrode potential. Its representative studies are as follows: (1) The constant‐potential charge equilibration simulation framework (DP‐QEq) separates long‐range from short‐range energy interactions and adopts a dynamic charge iteration algorithm to accurately describe long‐range electrostatic interactions at interfaces. It realizes constant‐potential boundaries based on the DFT‐QEq method. Combined with the dual‐electrode model, it can faithfully recreate the operational environment of batteries, and explain the internal mechanism that inorganic components of the SEI tend to trigger lithium accumulation and dendrite nucleation at the atomic scale. Nevertheless, this method produces minor errors in charge calculation, and the simulated voltage cannot fully match the actual operating conditions of batteries [123]. (2) Hybrid scheme of MLPs to treat electrochemical interfaces (ec‐MLP) overcomes the exorbitant computational overhead of traditional ab initio simulations. It can simultaneously characterize the interfacial charge and polarization features of both electrodes and electrolytes, efficiently reproduce the capacitance and dielectric distribution rules of electrochemical interfaces, and provide a new method for analyzing the microscopic mechanisms of complex electrode interfaces [124]. (3) Based on the Poisson–Fermi–Dirac equation and combined with DFT, the solid‐state electric double‐layer model can accurately describe band bending, defect evolution and ion‐electron coupling behaviors at solid‐state interfaces. It is able to optimize the thickness of interfacial layers and erode the migration barrier of lithium ions, providing quantitative design guidelines for the interface modification of all‐solid‐state batteries [125]. The charge transfer method adjusts the number of electrons by balancing with an external electron reservoir to maintain the system at a constant electrode potential. Its representative achievement is redox split charge equilibration (redoxSQE). On the basis of the standard split charge equilibration (SQE) method, it introduces the integer charge transfer (ICT) mechanism. Researchers can adjust parameters to observe the real‐time response of the system, directly and efficiently simulate electron transfer at the EEI, which is the core electrochemical process, pioneering a brand‐new computational route for the rational design of electrochemical energy storage devices [126]. The above methods break the simulation limitations of traditional MLFFs from various dimensions, and have initially established a simulation framework for describing the complex dynamic behaviors of charged interfaces. In potentiostatic electrochemical systems, the two methods are applied in combination. The charge equilibration method is used to simulate the behaviors of ideal polarizable electrodes and is good at describing charge accumulation and polarization on electrode surfaces, while the charge transfer method corresponds to the model of ideal nonpolarizable electrodes and focuses on interfacial electron transfer, that is, discrete transitions of integer charges [127]. Together, they cover part of the response spectrum of electrochemical interfaces driven by external potentials. More specifically, when studies focus on phenomena dominated by long‐range electrostatic effects such as electric double layer structure and ion accumulation, the charge equilibration method is the first choice due to its high computational efficiency and the capability to directly apply constant‐potential boundaries. When studies involve short‐range processes of material dissolution and deposition featured by discrete changes in atomic oxidation states, the charge transfer method is a proper option. In terms of dynamic process simulation, the charge equilibration method and the charge transfer method adopt different physical principles to realize the atomic‐scale description of electrochemical interface evolution, and have been successfully applied in several core scenarios:

1. Electrolyte decomposition [128, 129, 130]. Under constant‐voltage simulation conditions, the charge equilibration method can directly capture the full cascade of electrolyte molecule decomposition at interfaces and identify corresponding reaction pathways. Models of the charge transfer method do not involve chemical bond cleavage, so they cannot directly simulate the molecular decomposition of electrolytes. Nevertheless, the ICT mechanism can be used to evaluate the tendency of electrolyte components to undergo reduction reactions and explain from the electronic perspective which solvent molecules are more susceptible to decomposition.

2. SEI formation [131, 132]. Simulations such as DP‐QEq reveal the inhomogeneous composition distribution of SEI layers. This method verifies the internal correlation at the atomic scale: lithium atoms accumulate in advance in inorganic SEI components, induce uneven deposition, and eventually trigger dendrite nucleation. The growth positions of dendrites highly overlap spatially with the distribution of inorganic SEI components. Since models including redoxSQE are incapable of simulating chemical reactions, they cannot reproduce the formation of SEI films, but can likewise explore the key properties of SEI.

3. Interface evolution [133, 134]. Through four consecutive rounds of constant‐potential simulations (300 ps per round), DP‐QEq dynamically presents the entire evolution process of lithium dendrites, including nucleation, growth, tip exfoliation and the formation of isolated lithium clusters with a size of approximately 1 nm. By virtue of the ICT and ICTM mechanisms, redoxSQE enables the direct simulation of electrode dissolution and deposition: atoms on the anode surface lose charges, turn into ions and dissolve into the electrolyte; cations in the electrolyte migrate to the cathode surface, get lowered and deposited, forming uneven aggregates and dendritic structures. Therefore, the charge equilibration method is suitable for the direct observation of electrolyte decomposition, SEI formation and dynamic dendrite evolution under constant potential conditions, while the charge transfer method focuses on interpreting the underlying principles of electrode dissolution, deposition and morphological evolution. The two methods complement each other in the simulation of dynamic processes at electrochemical interfaces.

Even so, both methods have inherent limitations. The accuracy of the charge equilibration method highly relies on empirical electronegativity parameter sets, resulting in limited capability to describe strongly electron‐correlated systems and potential inaccuracies in characterizing local complex chemical environments. Even after ML acceleration, the charge transfer method still suffers from high computational cost, delivers insufficient accuracy in handling long‐range interactions, and its simulation time scale fails to cover the complete dynamic processes [123, 126]. The combined application of the two methods is expected to overcome their respective limitations via multi‐scale strategies. Spatially, the charge equilibration method is adopted to simulate ion distribution and polarization response in electric double layer regions, while the charge transfer method is used to analyze redox reaction kinetics within several nanometers of electrode surfaces. Temporally, ML acceleration connects picosecond‐scale electron transfer and nanosecond‐scale SEI evolution. This provides a feasible approach to reveal the microscopic mechanisms of dynamic reaction processes at the atomic scale. Looking forward, a comprehensive unified theoretical framework will be developed in the future to further guide the atomic‐scale simulation of interfaces.

3. Applications of ML in Li Electrode–Electrolyte Interface Engineering

3.1. Electrodes and Electrolytes Screening

The Li electrode–electrolyte interface reaction is affected by the collective influence of multiple factors, including electrodes and electrolytes compositions, electrolyte properties, and interfacial dynamics, which collectively determine battery performance. The high‐dimensional parameter space inherent to these systems renders conventional trial‐and‐error experimental screening and first‐principles computational approaches computationally prohibitive and statistically underpowered for targeted electrode or electrolyte design. Benefiting from the powerful analytic algorithms, ML can “learn” complex correlations and patterns from the existing data, providing insightful guidance for candidate exploration. High‐accuracy prediction and screening constitute the cornerstone of materials discovery and performance evaluation in advanced material systems. The initial phase entails data acquisition and feature generation, with raw data primarily sourced from established databases, experimental measurements, computations, simulations, and published literature. Subsequently, descriptors are constructed to encode key material attributes such as structure, composition, and properties. These descriptors are then processed to build curated datasets, which are used to train ML models. The trained models are finally leveraged for predictive performance analysis and materials discovery. Crucially, the reliability of ML‐driven predictions is contingent upon iterative experimental validation loops.

Leveraging powerful learning algorithms, ML has demonstrated its versatility in recent work by enabling the rapid discovery of interface‐stabilizing advanced materials, placing special emphasis on electrolyte engineering. For liquid electrolytes, interfacial ion transport is a key determinant of the reaction kinetics. Generally, properties of Li⁺‐solvent pairs serve as a critical parameter in the investigation of the ion transport phenomenon, and the screening of lithium salts and solvents for Li⁺‐solvent pair manipulation has emerged as a central strategy for electrolyte design [135]. A representative study was demonstrated by Chang et al. [79], a framework integrated synergistic computational–experimental with ML was developed to assist electrolyte screening (Figure 6a). A systematic MD campaign was executed across 166 distinct electrolyte formulations, comprising five lithium salts and eight ether‐based solvents, to quantify solvation thermodynamics and lithium‐ion diffusion coefficients (D_Li ⁺). Among seven candidate algorithms, GPR achieved superior predictive accuracy (R > 0.94), enabling reliable identification of nonintuitive composition trends. According to the prediction, 11 candidate electrolytes were identified, among which a ternary system with lithium bis(fluorosulfonyl)imide, 1,2‐diethoxyethane and lithium nonafluorobutanesulfonate was demonstrated as the optimal solution with the highest CE of 98.32% and the lowest overpotential.

FIGURE 6.

Schematic of electrode and electrolyte screening methods. (a) Integrated workflow combining MD simulation, ML, and experimental validation for electrolyte optimization. Reproduced with permission [82]. Copyright 2025, Wiley‐VCH GmbH. (b) The schematic uses database‐derived fingerprints to predict E_b . Blue/red elements mark octahedral (M1) and square‐planar (M2) sites. Features (N, D_p , R _cov, G_i , E_p , E_i ) represent atomic number, dipole polarizability, covalent radius, group index, Pauling electronegativity, and first ionization energy. Reproduced with permission [144]. Copyright 2025, Springer Nature. (c) Narrows dual‑doped LLZO candidates from 10 368 to 344 through ML screening. Reproduced with permission [145]. Copyright 2025, Elsevier B.V. All rights reserved. (d) Ion conducting properties of newly predicted versus known solid‐state Li‐ion conductors. Reproduced with permission [148]. Copyright 2019, Springer Nature. (e) A CGCNN and TL‐based framework for property prediction and screening of SSEs. Reproduced with permission [150]. Copyright 2025, Elsevier B.V. All rights reserved.

Besides, to achieve enhanced safety and higher energy density, shifting to all‐solid‐state LMBs with SSEs facilitates the advancement of such batteries in which unwanted chemical side reactions and dendrite penetration can be suppressed. To date, a variety of solid polymer electrolytes (SPEs) [136, 137, 138] based on polymer host (e.g., poly(methylmethacrylate), poly(ethylene oxide), poly(vinylidene fluoride‐co‐hexafluoropropylene)) and Li salts (e.g., LiTFSI), and inorganic solid electrolytes (e.g., garnet‐type (Li₇La₃Zr₂O₁₂), perovskite‐type, and sulfide electrolytes) [139, 140, 141, 142], as well as solid composite electrolyte (SCE) have been developed. Serving as the “blood” in LMBs, the solid electrolyte undertakes the key responsibility of conducting ions, dominating the output performance of battery [143]. Despite decades of research, limited solid Li⁺ conductors exhibit super ionic conductivity. Thus, this scarcity makes the discovery of new, highly conductive SSEs a defining challenge and a major breakthrough enabling commercialization of all‐solid‐state LMBs. ML techniques have developed a new paradigm for SSEs discovery. For instance, Lao et al. [144] employed ML‐driven modeling for fast screening and prediction of potential SPE (Figure 6b). In this investigation, 308 different candidates were generated to comprise the data input for ML analysis. Assisting with UL, a non‐monotonic (inverse volcano) relationship in ion diffusion barrier for M(DMF)₂Ni(CN)₄ (M  =  Mn²⁺, Fe²⁺, Co²⁺, Ni²⁺ or Cu²⁺) was mapped, identifying the Co(DMF)₂Ni(CN)₄ as optimal candidate with the highest ionic conductivity. In‐depth electronic structure analysis showed that the competitive binding affinity between transition metal–oxygen (M–O) and lithium–oxygen (Li–O) bonds served as a critical design principle for engineering a weakly coordinated Li⁺ solvation environment to facilitate ions migration. As demonstrated by Co‐Ni‐DMF, the strong M–O bonds tethered DMF locally and simultaneously weakened Li⁺ coordination. Consequently, the best electrochemical performance was achieved by SPEs‐based cells utilizing Co(DMF)₂Ni(CN)₄ among those with different Hofmann structures, achieving stable Li plating/stripping in Li–Li symmetric cell for 5000  h at 0.1  mA cm⁻² with an initial polarization of 27.9  mV, highlighting the great promise of ML‐driven approach in materials screening. Similarly, ML approaches to identify promising inorganic solid electrolytes are also getting more attention. As demonstrated in Figure 6c, 61 candidates were derived from 10 368 dual‐doped Li₇La₃Zr₂O₁₂ SSEs considering the characteristics of band gap, ionic conductivity and elastic properties via the combination of ML and DFT [145]. Subsequently, 5 potential candidates were functioned as fillers to fabricate poly(vinylidene fluoride‐co‐hexafluoropropylene) SCE. Encouragingly, the obtained SCE exhibited improved performance in comparison with that without fillers. Especially, the SCE with Zn 0.25 and Ti 0.25 mol dual‐doped Li₇La₃Zr₂O₁₂ filler demonstrated the optimal performance with an ionic conductivity of 8.7 × 10⁻⁴ S cm⁻¹ at 25°C, which endows the assembled solid‐state LMBs with enhanced cycling stability (capacity retention of 91% at 100 cycles) compared to the counterpart (capacity retention of 82% at 100 cycles).

Despite significant advancements in the field, the development of ML models for solid electrolytes faces great challenges compounded by the scarcity of experimental data and inherent challenges associated with high‐ or low‐dimensional feature spaces. These factors contribute to an elevated imbalance in both data and models, exacerbating the problem of overfitting [146, 147]. Fortunately, the emergence of comprehensive experimental materials databases, such as the MP, the ICSD, and the Cambridge Crystallographic Data Centre (CCDC), provides unlabeled data solution for SSEs exploration via the UL. For example, to discover candidate materials from Li‐containing compounds, Zhang et al. [148] implemented a two‐step workflow: unsupervised screening of ICSD entries yielded unlabeled modified X‐ray diffraction (mXRD) patterns, which were then clustered using limited conductivity data to prioritize promising candidates. As a result, using AIMD simulations, 16 fast Li‐conductors were screened, with predicted conductivities of 10⁻⁴−10⁻¹ S cm⁻¹ (Figure 6d). Although it has been shown UL can overcome the scarcity of labeled data to discover and characterize new materials, clustering for narrow down candidate material lists for further exploration does not establish a direct relationship between the structure and properties of the target materials. These models also fall short of effectively harnessing the information contained in existing labeled data and tackling the unlabeled data. By contrast, SSL methods have made substantial strides in boosting model performance under labeled‐data scarcity, tapping into copious unlabeled data. A representative study was demonstrated by Wan et al. [149], the SSL framework was performed to discover the promising fast Li‐ion conductors. This data‐centric strategy, which couples the ICSD with the use of unlabeled data to mitigate label scarcity, produced robust models (ensemble F1 score: 0.92) and led to the discovery of 38 superionic conductors. Of these, 32 align with known reports, and six new candidates were substantiated by ab initio simulations. Li₆AsSe₅I was subsequently synthesized, with experiments confirming its superionic conductivity. Besides, another critical challenge lies in the paucity of efficient data for specialized material properties (e.g., mechanical strength), which hinders the training of robust predictive models. TL addresses this by enabling a model pre‐trained on large‐scale related datasets to transfer knowledge to a target domain in low‐data regimes. Lateral transfer, a key branch of TL, involves knowledge transfer between analogous tasks, which is typically applicable under shared feature and distributional characteristics across domains. Pre‐training the model on the source task and transferring the acquired information to appropriately define the target task enables effective harnessing of pre‐existing knowledge to boost the performance of specific data‐scarce tasks, while significantly cutting down target task training time and data demands. As shown in Figure 6e, the work by Tao et al. [150] offers supporting evidence for this mechanism. In their study, 21 686 Li compounds were employed as input to train an initial model to predict their band gap, among which 889 structures in the data have known shear modulus (Gs) were utilized to further tune the model. Based on the CGCNN, the relationship between band gap and Gs was captured and the high Gs prediction accuracy of 90% was achieved. Ultimately, 12 promising SEI candidates with high Gs values (>27.2 GPa) and low diffusion energy barriers (<0.4 eV) were pointed out. The low diffusion energy barriers were beneficial to promote the ion diffusion and reduce the concentration polarization, thereby suppressing the dendrite formation. Besides, the high Gs endowed the interface with the capability of withstanding the stress generated during Li dendrite growth, thus mitigating dendrite propagation and improving the stability and safety of SSBs.

Furthermore, the challenge posed by the scarcity of datasets is further exacerbated for breakthrough technologies such as anode‐free LMBs. To handle noisy data effectively, a Bayesian Model Averaging (BMA) approach is developed, pooling multiple surrogate models to collectively generate predictions with quantified uncertainties and form a robust ensemble that balances the performance of individual models [151]. For example, to successfully screen for the ideal candidate materials identified despite typical constraints of data scarcity and label noise, Amanchukwu and co‐workers adopted a combined approach of sequential Bayesian experimental design with BMA [152]. By iteratively optimizing predictions with experimental feedback on Cu–LiFePO₄ capacity retention starting from 58 data points and exploring a million‐electrolyte space via AL, seven campaign rounds (∼10 tests per round) yielded four unique solvents that match the performance of leading benchmarks. The above research illustrates that AL methods have good application prospects in exploring the vast electrolyte chemical space required for future batteries.

In materials screening, structural information often proves decisive for performance outcomes. To accurately identify promising candidate materials, ML can be employed to map their interatomic and long‑range interactions, particularly in systems with characteristic structural motifs. Notably, this method strongly benefits scenarios with extensive database resources. Therefore, there is a demand to assemble a comprehensive and reliable database, and optimally, to consolidate multi‐sourced, heterogeneous, high‐dimensional, and sparse data to facilitate interface optimization.

3.2. Extracting Powerful Descriptors

Developing universal correlations between structure and property coupled with underlying mechanisms is a central objective of chemical research, essentially entailing the distillation and generalization of the underlying patterns from complex systems. Based on the established relationship, ML algorithms help extract efficient material descriptors derived from limited computational or experimental datasets. The obtained high‐precision descriptors can provide accurate input features for guiding the electrode and electrolyte design. Recent studies have identified multiple descriptors, which can be mainly categorized as electronic descriptors (e.g., dielectric constant, dipole moment, donor number, the LUMO energy level and the highest occupied molecular orbital (HOMO) energy level) [153, 154] and thermodynamic descriptors (e.g., binding energy (BE) to Li⁺ [155], electrostatic potential [156]) that correlate with the battery stability and CE.

Solvents as the main constituent of electrolytes are critical for modulating electrolyte properties, with solvent screening being a promising strategy for electrolyte design. To be specific, property of the Li⁺–solvent pairs has a profound effect on ion transport and surface adsorption/desorption behaviors and hence affects the deposition kinetics. Additionally, the continuous interfacial reactions between Li metal anodes and solvents can trigger electrolyte breakdown and electrode degradation, thereby leading to rapid capacity fade and accelerated cycle life decay. Therefore, various descriptors of the ion‐solvent complex have inspired new insights into electrolyte design and achieved significant success in LMBs. Since solvents in the solvation shell of Li ions are prone to decompose, leading to the instability of the electrolyte–anode interface, the LUMO that reflects reductive stability of solvents is widely demonstrated. As illustrated by Gao et al. [157], leveraging an innovative graph‑based method, a database of 1399 solvent molecules was developed, for subsequent analysis via integrated first‑principles and ML approaches. As illustrated in Figure 7a, 99% of the solvents delivered a reduced LUMO energy level after coordination with a Li ion. The LUMO energy change scaled positively with the BE, along with Li–O bond lengths and C–O bond distances, driven largely by the weight of carbon 2p orbitals in the LUMO. Moreover, analysis based on the RF algorithms revealed the fundamental physicochemical drivers of ion‐solvent complex stability by identifying the dipole moment (μ) and molecular radius (R) as the key descriptors for reductive stability. This insight significantly accelerates the rational screening of solvents. Besides, considering that thermodynamics of ion solvation in an aprotic electrolyte is primarily driven by strong enthalpic interactions between electron donor groups on the electrolyte components and Li ions, these electron donor interactions can be effectively described by the donor number (DN), which represents the Lewis basic character of the molecule [158]. Although various techniques allow to predict DN, the impact of DN on ion‐molecule binding and coordination, and its subsequent effect on solvation structure, remains underexplored. Moreover, challenges arise in obtaining accurate DN values through experimental testing, especially for the solid polymer electrolytes with the intricate nature of polymer conformations. Advances in ML technology have made it possible to leapfrog this difficulty. As investigated by Gao et al. [159], DFT calculations on the ion‐molecule intricate model demonstrated a strong linear correlation between DN and both the HOMO energy level and BE of ion‐molecule complexes, establishing DN as a predictive descriptor for Li⁺ coordination strength. Leveraging the ML‐driven generative framework shown in Figure 7b, a library of structurally diverse polymer monomers with different DN groups was designed. Molecular dynamics simulations and Raman spectroscopy revealed that polymer segments enriched in high‐DN groups were favorable to effectively reorganize the Li⁺ solvation shell, suppress ion aggregation, and promote homogeneous ion distribution. As a proof of concept, the resulting zwitterionic polymer electrolytes exhibit enhanced Li⁺ transport kinetics and prolonged cycling lifespan, validating DN‐guided molecular design as an efficient strategy for next‐generation solid electrolytes. In addition to the aforementioned traditional descriptors based on isolated molecules, dynamic descriptors that can capture the intrinsic correlations at the solvent level have emerged as a new research hotspot, addressing the limitations of single molecular descriptors in complex electrolyte systems. Kaviani et al. [160] established a database containing 103 types of ionic liquid@polyvinylidene fluoride (PVDF) gel polymer electrolytes by combining MD simulations and ML. They extracted multiple characteristic parameters, and conducted batch calculations of ionic conductivity as well as model training. The self‐diffusion coefficient of lithium ions was calculated via the mean square displacement (MSD), and the ionic conductivity was further derived based on the Nernst–Einstein equation. Analysis using the SHAP tool revealed that the Li⁺ diffusion coefficient was the core factor affecting conductivity, and showed a nonlinear correlation with the Pugh's ratio which characterized mechanical toughness: the faster the lithium ions migrate, the higher the conductivity. By contrast, moderate material toughness further enhanced this beneficial effect. Essentially, a higher diffusion coefficient meant stronger ionic conductivity, and a polymer matrix with appropriate flexibility could form unobstructed ion transport channels. The combination of the two achieved an optimal trade‐off between high conductivity and excellent robust mechanical performance. This research proposes evaluation indicators obtainable through simulations, together with an interpretable ML analysis framework, providing an effective method for the targeted development of such gel polymer electrolytes. Wang et al. [161] started from the solution structure and adopted the universal ML potential (uMLP) combined with parallel learning strategies to build an electrolyte database with over 2300 randomly matched solvent molecules. Relying on this database, they performed batch calculations and result verification of core properties including electrolyte density, viscosity, ionic conductivity and applicable temperature range. By calculating the integral of the autocorrelation function for each potential coordination bond pair around lithium ions, the researchers defined the average coordination lifetime (τ̅) as a quantitative criterion for solvation strength. The results verified that a shorter coordination lifetime of lithium ions corresponds to higher ionic conductivity of the electrolyte. Fundamentally, a short coordination lifetime indicated a weak solvation environment in the system, which reduced the desolvation energy barrier of lithium ions and thereby improved ionic conductivity. This body of work fills the gap in the rational screening of electrolytes. It not only puts forward quantitative evaluation indicators that can be directly acquired through simulation calculations, but also provides an efficient screening tool to facilitate the research and development of electrolytes with weak solvation and high ionic conductivity. Nevertheless, accurate quantification of solvating power remains challenging when relying solely on the single descriptor such as LUMO energy or the DN, both of which inadequately capture the complex interfacial dynamics governing composition and architecture of SEI. Although the LUMO energy level can describe the reductive tendency, the LUMO will undergo a downshift upon Li⁺ coordination, which strongly scales with HOMO energy level of pure solvent. It follows that the HOMO–LUMO gap that integrates both electron‐donating capacity and electron‐accepting properties is more deterministic [162]. However, it should be clarified that the HOMO–LUMO gap corresponds to the width of the electrochemical window of the electrolyte, namely the difference between the oxidation potential and the reduction potential. It cannot independently characterize the reductive decomposition tendency of solvents. Moreover, in real electrolyte systems, ion coordination and intermolecular interactions will cause a substantial shift in the energy levels of frontier molecular orbitals, which further limits the application effect of a single orbital parameter. The DN, traditionally defined for isolated solute molecules in weakly coordinating solvents, fails to reflect the solvation behavior in the electrolyte where ion–solvent interactions are highly cooperative and concentration‐dependent. And the descriptor of dipole moment is generally derived from the isolated models, which omits the influence of intermolecular polarization [163]. In terms of the BE, it typically accounts for the binding of one Li⁺ ion to an individual solvent molecule, while neglecting the coordination of multiple solvent molecules with Li⁺ and the competitive solvent–solvent interaction behavior. The mechanistic research recently conducted by Li et al. [164] further highlighted this deficiency: by combining nuclear magnetic resonance and Raman spectroscopy, the study systematically demonstrated that relying solely on the binding energy of isolated ion‐solvent pairs or the LUMO energy level of single molecules fails to accurately characterize electrolyte performance and interfacial behaviors. The results revealed that chemical shift displacement of DME occurs in the DME/DOL composite system, which was not induced by variations in the intrinsic binding energy between Li⁺ and DME, but by dynamic dipole–dipole interactions between electron‐rich oxygen (δ⁻) in DOL molecules and electron‐deficient hydrogen (δ⁺) in DME molecules. Such dynamic solvent–solvent interactions directly weakened the coordination capacity between Li⁺ and DME, drastically reduced the desolvation energy barrier and facilitate lithium deposition. Nevertheless, static isolated molecular descriptors could not fully capture this interaction mechanism. To overcome this limitation, an interpretable descriptor based on hydrogen‐bond acceptor basicity (β) and HOMO–LUMO gap was generated by ML‐based method from a set of feature inputs, including β, DN, dipole moment, Li chemical shift, BE of a solvent with a Li⁺, LUMO and HOMO energy levels, and the HOMO–LUMO gap [153]. Herein, β corresponds to Li⁺ solvating ability of the solvent, governing the occurrence probability of solvent molecules on Li metal anodes, and the HOMO–LUMO gap can reflect the overall electrochemical stability range of the system. Combined with the basicity of hydrogen bond acceptors, it can comprehensively evaluate the overall interfacial stability. The integration of β and HOMO–LUMO gap determines the stable and uniform SEI formation, thereby affecting the battery CE. Moreover, the established quantitative regression model was established that correlates the generated descriptor with the CE, as shown in Figure 7c. The results indicated that the solvents with a moderate β _normalized (0.45–0.65) and a high HOMO–LUMO gap (>16.9 eV) were the promising candidates to achieve high CE. As a proof of concept, a novel solvent ethylene glycol ethyl methyl ether was identified, which endows high CE for both the coin cell (99.8%) and the pouch cell (99.2%), significantly facilitating the search for advanced electrolytes.

FIGURE 7.

Schematic diagrams of electrolyte design research correlating microscopic descriptors with macroscopic performance. (a) Variation in LUMO energy level versus carbon 2p orbital contribution ratio. Reproduced with permission [157]. Copyright 2023, American Chemical Society. (b) Integrated workflow diagram for DN prediction and application. Reproduced with permission [159]. Copyright 2024, Wiley‐VCH GmbH. (c) Regression model for estimating CE based on β and HOMO–LUMO energy gap. Reproduced with permission [153]. Copyright 2025, Wiley‐VCH GmbH. (d) SEI features from cryo‑TEM. Reproduced with permission [166]. Copyright 2025, Springer Nature. (e) Causal DAG graph constructed based on LiNGAM, with experimental electrode potential E _exp defined as the outcome variable and only connections with absolute coefficient values ≥ 1.0 retained (n30). Reproduced with permission [167]. Copyright 2025, American Chemical Society. (f) Schematic of bimodal COSNet integrating dual GCNs, attention and MLP for material property prediction. Reproduced with permission [169]. Copyright 2024, American Chemical Society. (g) Experimental conductivity verification of generated electrolyte formulations with defined target values and solvent abbreviations. Reproduced with permission [170]. Copyright 2026, Springer Nature.

The substitution of conventional liquid electrolytes with solid ones serves as a powerful approach to mitigate uncontrolled dendrite formation and the associated explosion risk. Advances in ML technology have emerged as a transformative approach to mitigate the persistent challenges associated with solid electrolytes, including the poor ionic conductivity and low Li⁺ transport numbers. Taking the solid‐state composite polymer electrolyte as an example, regulating lithium salt dissociation kinetics by strengthening the interaction between inorganic fillers and lithium salts is vital for enhancing its ionic conductivity. HT DFT calculations and ML algorithms extracted a novel descriptor, λ(d/p) that reflects d–p orbital coupling strength, which is validated powerful to describe the energy barrier for lithium salt dissociation [165]. Based on the established descriptor, Ti atoms with the optimal d–p orbital coupling strength were identified as a promising element to balance lithium salt dissociation to realize the improved ion mobility. However, it should be mentioned that both too weak (e.g., Mn, Fe, Co, and Ni) and excessively strong d–p coupling (Sc, V, and Cr) resulted in limited ion‐pair separation. As validated by experiments and simulations, electrolytes with ZIF‐8@Ti exhibited the enhanced ion dissociation kinetics and obtained an appreciable ionic conductivity, reaching 10⁻³ S cm⁻¹ at 30°C. Apart from the significant role of enhancing Li⁺ transport, electrolyte engineering has been developed to tailor the composition and spatial distribution of inorganic (e.g., LiF, Li₂O, Li₃N) and organic species within the SEI layer, thereby enabling homogeneous lithium deposition through enhanced interfacial stability and mechanical integrity. Nevertheless, the general correlation between SEI composition and Li deposition morphology remains poorly understood. Multiscale nonlinear interactions between these two factors, coupled with the limited capacity of state‐of‐the‐art characterization techniques to analyze large‐scale, heterogeneous, and high‐dimensional datasets, render the accurate understanding and prediction of Li deposition behavior a formidable challenge to date. To overcome such limitations, the concept of SEI‐omics was developed to demonstrate the correlations between SEI properties and Li single crystal structures [166]. In the exploration workflow, an extensive cryo‐TEM experimental dataset was constructed as input. Subsequently, a ML data‐driven approach was employed to extract the determinate feature. This process is shown in Figure 7d. The geometric structure factor, λ, defined as the axial‐to‐radial growth ratio, provides a metric for quantifying the nanoscale architecture of Li metal nanocrystals. Based on the value of λ, Li deposition can be classified into dendritic Li (λ > 1) and spherical Li (λ ≤ 1), which provided a quantitative framework to directly correlate specific SEI compositions and structures with distinct lithium deposition mechanisms. Conclusions are drawn from the analysis of a series of SEI layers combined with multiscale simulations. The elevated surface energy and ionic mobility of the SEI layer synergistically promote the spherical Li deposition. An excessively disordered SEI was achieved through ML‐guided composition design, leading to a uniform lithium deposition characterized by an exceptionally low λ value (approximately 0.65), and endowing the Li–Cu cell with a high average CE of 99.35% over 800 cycles at 1  mA cm⁻² and 1  mAh cm⁻².

It can be seen that ML offers a robust framework to reveal the hidden mechanisms and powerful descriptors among the complex components of the electrolytes and SEI, and guide the design of new electrolyte systems and advanced SEI to stabilize the EEI. However, further analysis reveals that constrained by theoretical assumptions and applicable scopes, numerous existing descriptors deliver unsatisfactory practical performance. This conclusion can be explained and verified by the widely adopted composite β‐ΔE descriptor available today. The combined descriptor (β‐ΔE), established by integrating the hydrogen bond acceptor basicity (β) and frontier molecular orbital energy gap (ΔE = LUMO–HOMO) proposed by Zheng et al. [153], effectively correlated the intrinsic properties of solvents with the Coulombic efficiency of batteries, providing a reliable basis for the rational screening of conventional electrolyte solvents. It should be noted that the model training and rule verification of this combined descriptor were all completed using datasets of conventional aprotic solvents such as ethers and carbonates, and it exhibits excellent predictive capability and interpretability in traditional electrolyte systems. Nevertheless, this descriptor has a clear scope of application. For special electrolyte systems such as highly fluorinated solvents and ionic liquids, the solvation structures and interfacial reduction reaction mechanisms derived from strong electron‐withdrawing groups or ions differ significantly from those of conventional systems. As characterization parameters for the properties of isolated molecules, β and ΔE see a sharp decline in prediction reliability. As shown in Figure 7e, the causal analysis conducted by Yokoyama et al. [167] offered crucial theoretical support for this finding: in concentrated electrolyte systems, intermolecular interaction parameters including the spatial distribution of lithium ions and anions as well as long‐range integral functions exhibited a direct causal relationship with the measured potential shift of electrodes, while intrinsic molecular orbital parameters such as HOMO and LUMO no longer played a dominant role in strongly correlated systems. This conclusion directly proves that the β − ΔE combined descriptor built on the properties of isolated molecules cannot be extrapolated to complex solvent systems dominated by strong interactions. To address this limitation, it is recommended to calibrate the model with a small amount of experimental data when applying it across solvent categories, such as extending from conventional carbonate systems to highly fluorinated solvent or ionic liquid systems. Meanwhile, auxiliary descriptors including solvation free energy and coordination enthalpy can be adopted for model optimization to broaden the applicable scenarios of this method. Overall, this work clarifies the applicable scope of the descriptor, reveals the physical causes of its failure, and proposes targeted expansion strategies.

Notably, the current research paradigm of electrolyte informatics is undergoing a profound shift from molecular descriptor‐dominated approaches to multimodal data‐driven models. Multimodal data fusion is not a simple combination of different information, but the alignment and joint learning of information from diverse sources and scales, such as spectral experimental data, MD trajectories, and DFT calculation results, to offset the biases of single‐modal data. Starting with external multimodal literature data mining, Lee et al. [168] adopted a solution combining LLMs and the automatic chart analysis tool MatGD. They performed text mining on 3606 Elsevier papers using GPT‐4, an LLM, and extracted 15 398 sets of battery composition data and 10 242 sets of battery cycling performance data. By matching text and graphical data, they established a standardized battery database containing 8074 entries covering composition, test conditions and capacity. Based on this database, they built models to predict the initial capacity and the capacity after 50 cycles of batteries. With the help of SHAP interpretability analysis, two key evaluation indicators for solvents were identified: the topological polar surface area (EState VSA6) and the molecular structural complexity (Kappa3). The research results proved that lower values of these two indicators correspond to higher initial capacity and better cycling stability of electrolytes. Further analysis demonstrated that electrolytes with low polarity and weak solvation capability facilitated the formation of anion‐derived SEI films and enable uniform crystallization and deposition of lithium metal, thereby extending the cycle life of batteries. This study addresses the long‐standing challenge of automatically integrating textual and graphical data from literature, puts forward quantitative evaluation indicators that can be directly calculated from molecular structures, and provides an integrated data mining pipeline for text and graphical information. The Wang team [169] further explored the internal material characteristics through multimodal data fusion and developed COSNet, a multimodal ML framework. As shown in Figure 7f, this framework integrated chemical composition and crystal structure data of materials, tackling the common problem that experimental data of solid electrolytes usually lack crystal structure information. They sorted out 1678 sets of lithium‐ion conductivity datapoints from 206 research papers, with conductivity values ranging from 10⁻¹² to 10⁻² S/cm, among which only 460 sets were accompanied by crystal structure data, demonstrating the necessity of multimodal learning. Leveraging the COSNet architecture, the framework not only predicted lithium‐ion conductivity but also simultaneously calculated multiple physical properties such as band gap and enthalpy of formation. Compared with conventional models that only utilize composition data without structural information, COSNet remedied training deficiencies via data augmentation, and its prediction accuracy was effectively improved through cross‐modal learning. This achievement solves the industry‐wide dilemma of scarce structural data for solid electrolytes, and presents an integrated solution combining transfer learning, ensemble learning and HT screening. It has facilitated the discovery of a variety of new fast lithium‐ion conducting solid electrolytes including Li_1.5Nb_0.5F_5.5. Breaking away from simple data and feature fusion, Yang et al. [170] constructed a unified framework that integrated performance prediction and inverse electrolyte formulation generation. This framework incorporated massive datasets of single molecules, experimental results and MD simulations, combined forward prediction models and inverse generation methods, and established a complete closed‐loop AI design system for electrolytes. A three‐stage hierarchical training strategy was adopted. The prediction module integrated GNNs and classic physical equations, achieving determination coefficients as high as 0.985 and 0.953 for the prediction of ionic conductivity and anion ratio respectively, as demonstrated in Figure 7g. For the generation module, a generative framework tailored for molecular mixture design was developed for the first time. Trained on 30 000 sets of formulas generated by the prediction model, it could satisfy multiple parameter constraints such as conductivity and anion proportion, enabling the targeted design of multi‐component electrolyte formulas. Experimental verification showed that 10 generated formulas meet the performance requirements, and three of them exhibited a conductivity of 15 mS cm⁻¹, vastly outperforming single‐solvent benchmark electrolytes. Through multi‐stage training and embedded physical rules, this framework efficiently explored a huge number of formulas on the scale of 1.65 × 10¹⁷, making the multi‐objective and multi‐constraint inverse design of electrolytes feasible. This work realizes the full workflow from performance prediction to autonomous formula design. All relevant models and datasets have been open‐sourced, which can serve as references for other mixed systems such as fuels and polymer blends.

Collectively, these studies demonstrate that multimodal data fusion has moved from conceptual exploration to practical application, and is playing a tangible role in the research and development of electrolytes and battery materials. Importantly traditional standalone molecular descriptors and emerging solution‐level data are not mutually exclusive, but complementary. The β − ΔE descriptors based on isolated molecules still feature high efficiency and strong interpretability in the preliminary screening of conventional solvents. For the refined evaluation of candidate solvents and the research on special electrolyte systems, hybrid descriptor models based on solution transport properties, solution structures and multimodal experimental data are required. This hierarchical descriptor strategy is expected to become the future development direction of electrolyte informatics, and also provides a more practical and theoretically in‐depth quantitative framework for the targeted design of electrolytes for high‐energy‐density lithium metal batteries.

3.3. Mechanism Elucidation

Lithium dendrite growth during cyclic operation of LMBs constitutes a fundamental degradation mechanism that critically compromises cycle life, CE, and energy density, thereby impeding practical deployment. Notably, when dendrite protrusions exceed the critical threshold of separator integrity, they can pierce through the separator and establish direct electrical contact with the cathode, triggering internal short circuits and posing severe safety hazards, including thermal runaway and combustion. A series of advanced experimental methods, including transmission electron microscopy (TEM) and operando X‐ray µCT can examine the micro‐morphology of lithium dendrites [171, 172]. However, these methods remain fundamentally limited by their spatial and temporal resolution, rendering the real‐time capture of nucleation events, growth kinetics, and interfacial dynamics during electrochemical cycling. To overcome these observational constraints, computational modeling has emerged as an indispensable complement, offering mechanistic insights into lithium deposition processes across multiple length and time scales. Established approaches include ab initio calculations for electronic structure analysis, reactive force field molecular dynamics for reactive surface chemistry, Monte Carlo simulations for stochastic deposition patterns, and ML‐enhanced molecular dynamics. Among these, ML techniques trained on quantum chemical data achieve an optimal balance between predictive accuracy and computational efficiency, enabling large‐scale simulations of dendrite formation with atomic‐level fidelity.

ML force fields, as shown in Figure 8a, trained on high‐accuracy datasets from small‐scale systems, have been demonstrated to accurately extrapolate to mesoscopic and macroscopic length scales among alkali metal systems and beyond, enabling atomistic simulations of complex electrochemical reactions. Crucially, the data‐driven formulation of the PES permits the integration of multiple DFT training sets, particularly those incorporating lithium atoms under working electrolyte environment conditions, thereby capturing the lithium dendrite evolution. As demonstrated by Zhang et al. [172], a MLFF framework incorporating an implicit solvation model was developed to accurately capture solvation effects in lithium electrolytes, enabling atomic‐scale simulation of dendrite growth dynamics. By decomposing DFT total energies into atomic contributions, an end‐to‐end ML model was trained and integrated AL with rigorous verification, achieving scalable and reliable prediction of in situ lithium morphology evolution. It was observed that the morphological transformation of lithium dendrites proceeded through two distinct stages. In the initial stage, driven by surface energy, the upper crystalline domains formed with distinct crystal phases or orientations from the lower ones, thereby generating grain boundaries. Subsequently, the upper crystalline domains migrated along the grain boundaries, simultaneously repairing the initial grain boundaries and achieving atomic coherence at the interface, which ultimately resulted in the formation of a kink morphology. Kinetic simulations reveal that dendrites tend to collapse rapidly at high temperatures, with no formation of crystalline domains or grain boundaries. The absence of the grain boundary pinning effect accelerated the surface diffusion of lithium atoms, a process that effectively inhibits the further dendrite growth. Accordingly, three key energy factors were extracted, as follows: (1) surface energy: the full evolution of morphology is ultimately driven by the high surface energy inherent in the starting geometry, (2) internal domain energy: during the morphological transformation, individual domains remain largely unaltered, (3) grain boundary energy: the grain boundary energy is the main reason for the change in the statistical value of bulk energy. The appearance of multiple domains and grain boundaries was observed in MLFF‐MD trajectories of different initial configurations at room temperature, illustrating that this was a common characteristic of lithium metal in the solvent environment. By generating energy barriers through blocking or anchoring effects, it is expected that multiple crystal domains and grain boundaries can inhibit the dendrites from collapsing further. Therefore, inhibiting the formation of such microstructural features may represent a viable strategy to prevent the emergence of stable, persistent dendrites. As validated, constructing the glassy lithium without multiple domains has been demonstrated as an effective approach to preventing dendrite formation [173]. Relevant simulation studies have demonstrated that amorphous glassy lithium can effectively eliminate the directional growth of lithium dendrites caused by grain boundaries, thereby enabling uniform deposition of lithium metal. It serves as a good solution to enhance the stability of lithium anodes. Currently, there are three mainstream methods for engineering glassy lithium: low‐temperature deposition regulation, electrochemical alloying, and interfacial SEI induction. Specifically, electrochemical deposition at low temperatures and high current densities can disrupt the regular arrangement of lithium atoms, inhibit the ordered growth of grains, and form an amorphous lithium structure. Experimental characterizations confirm that appropriately increasing the deposition rate can effectively induce the formation of amorphous structures. Lithium anodes fabricated by this method feature uniform deposition and excellent cycling reversibility, with performance far superior to conventional polycrystalline lithium anodes [173]. Electrochemical alloying leverages the electrochemically driven amorphization mechanism. During lithium insertion at room temperature, alloy anodes such as Li–Si, Li–Sn, and Li–Al in situ form amorphous lithium alloy structures. Rapid lithium insertion destroys the original lattice structure of materials and generates stable glassy phases. This technology is compatible with various high‐energy alloy anodes and enables precise regulation of the content and distribution of amorphous structures. The in situ formed glassy lithium alloys can greatly improve the interfacial mechanical stability, offering a new approach to restraining lithium dendrite growth [174]. Apart from this, optimizing the composition of the SEI to induce the formation of disordered structures on electrode interfaces has emerged as a popular modification strategy in recent years. Based on the multi‐dimensional analytical methods of SEI omics, researchers have confirmed that raising the contents of fluorine, nitrogen, sulfur and phosphorus while reducing the oxygen content in SEI films facilitates the construction of disordered interfaces. By preparing SEI films rich in amorphous lithium fluoride and lithium nitride, a glass‐like thin layer of 5–10 nm can be formed on the surface of lithium metal. This layer reduces the diffusion resistance of lithium atoms, promotes uniform nucleation and growth of lithium, and ultimately achieves nearly spherical and uniform lithium deposition with a sphericity coefficient λ of approximately 0.65 [166]. Although the above methods can yield glassy lithium structures partially or entirely on lithium anodes, the practical application of glassy lithium anode systems still faces multiple critical technical bottlenecks at this stage. While glassy solid electrolytes can realize dendrite‐free lithium deposition and effectively optimize the deposition morphology of anodes, crystallographic defects such as grain boundaries inherent to traditional crystalline lithium still trigger uneven lithium nucleation and accelerate the performance degradation of anodes [175]. More critically, glassy lithium is a metastable structure. It will spontaneously recrystallize, regenerating grain boundaries during long‐term battery cycling, which drastically weakens its capability to suppress dendrites. By contrast, academic research on the interfacial evolution rules and long‐term stabilization mechanisms of glassy lithium under actual operating conditions is still insufficient. Even if glassy lithium structures can be in situ fabricated via solid electrolyte thin films, there is no clear and comprehensive understanding of their structural evolution behaviors, interfacial side reactions and long‐term structural stability in real working environments [176]. In terms of preparation processes, there are prominent technical barriers to the precise and controllable synthesis of high‐purity bulk glassy metallic lithium. Most current studies realize the in situ deposition of lithium on copper substrates and solid electrolyte thin films, making it difficult to stably fabricate large‐sized and high‐purity bulk glassy lithium. Furthermore, practical battery applications impose stringent requirements on performance indicators including deposition current density and cumulative cycling capacity. Meanwhile, the high cost of electrolytes suitable for stabilizing lithium anodes further hinders the industrial implementation of this technology [177]. Therefore, while glassy lithium represents an attractive candidate for developing highly stable lithium metal anodes, its development is currently restricted by multiple issues including phase transition of metastable structures, immature preparation techniques and inadequate research on interfacial mechanisms under operating conditions. Further research is required to tackle these challenges one by one.

FIGURE 8.

ML‐based elucidation of transport and interface mechanisms in battery materials. (a) Architecture of MLFF Models. The local environment within the cutoff radius R _cut corresponding to different atom i is described by symmetry function values G for energy determination. Reproduced with permission [172]. Copyright 2022, Wiley‐VCH GmbH. (b) Automated dataset generation workflow (along ConstQ and ConstP trajectories). (c) Lithium charge state distribution in cyclic ConstP MD simulations (units in # of |e|). Reproduced with permission [123]. Copyright 2025, Springer Nature. (d) Cascading Li⁺ hop: red ion leaves an AlCl₄ ⁻ cage, followed by blue ion entering it (representative trajectory). (e) The proportion of cascading transitions, P _casc (number of cascades / total number of transitions), varies with the composition (x). Reproduced with permission [192]. Copyright 2025, Wiley‐VCH GmbH. (f) Comparison of grain boundary model schematics: Model‐GB (Li|LLZO GB|LZO) and Model‐A (Li|LLZO‐A|LZO). Reproduced with permission [193]. Copyright 2025, Springer Nature.

Besides various difficulties in the preparation of glassy lithium, the MLFF framework itself also has fundamental limitations. Despite exhibiting great prospects in analyzing the dendrite formation mechanism, its predictive accuracy in modeling Li electrodeposition and dissolution processes remains constrained by the inadequate representation of atomic charge variability. This limitation stems from the absence of a physically grounded treatment of solvation‐induced polarization effects on coulombic interactions, which are critical for capturing the dynamic charge redistribution at the electrode–electrolyte interface. To deal with such limitation, Hu et al. [123] developed a framework integrating MLFF with the charge equilibration (QEq) method to perform constant potential (ConstP) MD simulations for visualizing the kinetics of dendrites nucleation at interfaces of Li metal anode. As shown in Figure 8b, this framework enabled analysis of variable atomic charges within a MLFF framework. Additionally, this strategy exhibited quasi‐linear computational scaling with system size, substantially outperforming conventional QEq solvers in memory efficiency and parallelizability. Encouragingly, stable simulations of systems comprising thousands of atoms, incorporating long‐range electrostatic interactions were demonstrated. Moreover, the framework took the voltage between the electrodes into account, thereby providing a more realistic description for lithium dendrite formation. As a result, simulations demonstrated that heterogeneous lithium deposition is preceded by the accumulation of lithium within the SEI, serving as a critical initiator for dendritic nucleation. To be specific, during the initial electrochemical cycle, lithium atoms at the anode interface undergo oxidation and react with the electrolyte, leading to the formation of a nascent SEI layer composed of inorganic and organic species. Upon reversing the applied potential, lithium aggregated within the SEI matrix, which was visualized as localized high‐density clusters. These nonuniform Li accumulations amplified the inhomogeneous electric field distribution, initiating dendritic nucleation at sites of elevated interfacial curvature. As cycling progressed, dendrites evolved into elongated, needle‐like morphologies with increasing aspect ratios, culminating in tip fragmentation during the fourth cycle. This process generated isolated lithium clusters approximately 1 nm distant from the primary dendrite shaft, a phenomenon indicative of mechanical instability and potential electrical isolation. The entire process was shown in Figure 8c. The spatial correlation between charge distribution and SEI composition, as revealed by simulations at the final snapshot, further elucidated the mechanistic link between interfacial heterogeneity and nucleation sites. Inorganic SEI constituents (Li₂CO₃, Li₂O, and LiF) are preferentially enriched in the inner region adjacent to the electrode, forming a dense, ionically conductive barrier. In contrast, gaseous decomposition products exhibit diffuse, nonuniform distribution throughout the outer SEI.

According to the above‐mentioned analysis, the formed SEI layer at the EEI serves as a key factor in the performance of LMBs [20]. Typically, the SEI formation is attributed to the decomposition of the electrolyte. While the general origin of SEI is well established, the fine structure, chemical composition, microtopography, and formation mechanism of the SEI film have not yet been clarified so far, and its growth mechanism, in particular, remains inconclusive. Cutting‐edge analytical techniques, such as Cryo‐TEM, nuclear magnetic resonance spectroscopy, time‐of‐flight secondary‐ion mass spectrometry, and reflection ion interference microscopy, have enabled unprecedented molecular‐level resolution of SEI composition and spatial distribution [166, 178, 179]. However, limited by the extreme air‐ and moisture‐sensitive features and specific locations of SEI, these technologies are difficult to monitor the composition and structure evolution of SEI layer during the long‐term cycling. Many theoretical calculations elucidated the thermodynamic driving forces and kinetic barriers for SEI formation at the quantum chemistry level. For example, reactive force field molecular simulations captured the SEI growth on nanometer scales [180]. However, these approaches remain fundamentally limited by timescale constraints that are typically restricted to the very initial stage of SEI formation. Innovations in computational approaches are necessitated to better illustrate the genesis and evolution of SEI layers that span multiple length and time scales.

Recent advances in deep potential molecular dynamics (DeePMD) simulation show great potential to overcome this bottleneck, allowing for significantly larger model scales and longer simulation times without compromising the high accuracy and efficiency. This paradigm shift opens up new possibilities to simulate the multi‐stage SEI evolution. It has been reported that ML‐potential‐assisted MD simulations detected the SEI formation and growth at Li–β‐Li₃PS₄ interface based on the massive supercell containing 12 160 atoms corresponding to ∼40 nm [181]. According to the results, a layer with ∼12.4 nm formed between Li and β‐Li₃PS₄, which was composed of a Li₂S crystalline region with thickness ∼5.0 nm adjacent to Li and an amorphous one near the β‐Li₃PS₄ electrolyte. The SEI formation was demonstrated to undergo a four‐stage evolution process including (1) fast ion diffusion, (2) nucleation, (3) Li₂S growth, and (4) stabilization. Moreover, potential energy distribution of lithium showed that there was a significant drop of 1.42 eV over a 2.4 nm region, which functioned as a thermodynamic driver for parasitic reactions and interfacial decomposition. Subsequently, this primary energy barrier was partitioned into two distinct, lower‐magnitude gradients across the interfacial architecture with a 0.67 eV reduction at the SEI–β‐Li₃PS₄ interface (2.4 nm) and a 0.63 eV drop at the Li–SEI interface (1.8 nm). These stepwise energy transitions reflected the spatial evolution of charge redistribution and localized electric field screening as Li⁺ ions traverse the multi‐layered interphase. Apart from the in‐depth analyses of dendrite formation and SEI growth, a molecular‐level understanding about the mechanistic interplay between interface reactions and the Li dendrite growth is also critical. Recent investigations based on the DeePMD simulations showed that an isolated Li cluster in fact forms inside the SEI region, approximately 1 nm away from the Li anode/SEI interfacial boundary, rather than directly growing connected to the Li anode surface or within the SE bulk [182]. Moreover, it was pointed that a narrowed band gap in interstitial regions between lithium clustering sites and the anode surface promotes unintended electronic conduction, enabling in situ reduction of Li⁺ ions within the SEI matrix and thereby triggering localized lithium deposition. To suppress dendritic growth in all‐solid‐state LMBs, strategic engineering of the SEI's electronic structure to elevate its band gap is essential, to achieve electronic insulation and high ionic transference number. Furthermore, multi‐scale simulation techniques combining ML‐driven MD and phase‐field modeling have been employed to investigate the crucial role of charge transfer process on Li dendrite growth and the interface reactions [183]. It was found that the accelerated charge transfer kinetics at the initial stage of cycling enable efficient salt decomposition and the in‐situ generation of a compact, ionically conductive SEI on the lithium metal anode. This enhanced interfacial reactivity facilitated autonomous repair of intrinsic and cycle‐induced defects, thereby promoting the evolution of a morphologically uniform and mechanically robust SEI. Consequently, the stabilized interface enabled homogeneous lithium nucleation and reversible stripping/plating, effectively suppressing dendritic growth and enabling dendrite‐free operation.

The overall electrochemical performance of lithium metal electrodes is highly related to the ion transport behavior at the EEI. Understanding the ion diffusion mechanism is imperative for interface design. It is widely acknowledged that ions move through the liquid medium and electrons are separated in the liquid electrolyte. This dynamic environment facilitates ion transport and accelerates side reactions. Differently, ion diffusion occurs along energetically favorable pathways within crystalline lattices or traverses a disordered structure in the solid electrolyte, which eliminates leakage and improves the battery safety, yet exhibiting a relatively low ionic conductivity. Therefore, considerable efforts have been directed toward developing next‐generation SSEs with improved ionic conductivity toward the commercialization of all‐solid‐state batteries. To better guide the SSE design, understanding the intricacies of ion conduction is critical for optimizing battery performance. Nevertheless, revealing specific factors affecting ionic conductivity in SSEs remains challenging. From an experimental point of view, it remains challenging to apportion the relative contributions of bulk, grain‐boundary, and solid EEI conduction to the overall Li⁺ transport, either qualitatively or quantitatively. Furthermore, the complex and varied compositions and structures reported in experimental studies make it difficult to establish a consistent benchmark for lithium ionic conductivity. These challenges necessitate the development of comprehensive theoretical methods in order to elucidate and forecast the complex behavior of SSEs. Encouragingly, MD simulation based on the MLIP is a very powerful tool to unveil ion transport mechanisms in SSE due to the extended simulation time scale from ps level to microsecond (µs) level and high accuracy at the DFT level [184]. A variety of SSEs have been investigated through the ML‐based MD simulation, such as sulfide‐based inorganic SEs [185], lithium argyrodite [186], halide lithium SE [187], perovskite SEs [188], garnet‐type SE [189], highlighting the important role of local environment around Li⁺, atomistic structures, lattice dynamics and vacancies in facilitating ion diffusion for SSEs. Taking the Li₇La₇Zr₂O₁₂‐type perovskite SE as an example [188], the ML‐driven MD simulations revealed that dynamic distortions of the lattice lowered the activation barrier for Li⁺ hopping, thereby facilitating local ion mobility. Conversely, the local distortions of TiO₆ octahedra impeded long‐range ion diffusion. Drawing on these mechanistic insights, three design principles for high‐conductivity LLZO electrolytes were established: (1) expanding the percolation network of Li⁺ migration pathways, (2) promoting the dynamic lattice deformation of O4 square, (3) minimizing local octahedral distortions. As validated, a designed novel structural model featuring a linear alignment of three La³⁺ ions in every La‐depleted layer to suppress the distortions of the coordination octahedra endowed the SSE with a calculated ionic conductivity of 6.83 × 10⁻³ S cm⁻¹ at 300 K. Besides, MLFF‐based MD simulations were combined with neutron diffraction to further correlate the crystal structure with the ionic conductivity [190]. The neutron diffraction characterization revealed a nonstoichiometric Li distribution in the perovskite lattice, with 2.625 Li vacancies per unit cell and the residual 0.375 Li⁺ ions preferentially occupying the unconventional 24k Wyckoff site, which was distinct from the conventional A‐site (1a) in ideal perovskites. This anomalous occupancy pattern generated a percolating network of vacancy‐mediated migration channels, enabling anisotropic Li⁺ transport along interconnected pathways rather than isotropic hopping. However, MLFF‐based MD simulations illustrated that Li⁺ at the 24k site showed restricted mobility under low‐temperature conditions because of the strong Li–vacancy ordering that elevated activation energy barriers and reduced the ionic conductivity. These findings establish a direct structure–property linkage: precise control over cationic site occupancy and suppression of static vacancy ordering are critical design criteria for achieving high‐performance SSEs.

Compared to ordered structures, the absence of a well‐defined crystal lattice in disordered or glassy solid electrolytes hampers the accurate prediction and control of ion transport pathways [191]. The complex defective and disordered structure further obscures the accurate elucidation of ion transport mechanisms in such materials. Given the high computational demands of simulating diffusion processes at macroscopic length and time scales, in particular, at relatively low temperatures, MLFF simulations are widely adopted to model the ion diffusion in amorphous SE. As reported by Saito et al. [192], ML‐based MD simulations on inorganic glass SSE of Li _x AlCl_{3+ x} (x = 0.25 to 3) further confirmed the non‐monotonic composition dependence of Li⁺ conductivity as experimentally observed, with a maximum ionic conductivity at x = 1. This non‐monotonic composition dependence was mainly attributed to the interplay between local Li⁺ occupancy and vacancy concentration. Moreover, simulations presented in Figure 8d,e demonstrated that Li⁺ transport was predominantly governed by cascading hopping events, correlated jumps across adjacent vacant sites, rather than by paddlewheel mechanisms involving anion rotation. These cascading hops exhibited strong compositional sensitivity, directly underpinning the non‐monotonic conductivity profile. Furthermore, ML‐driven MD simulations were employed to investigate the role of grain boundaries on lithium segregation and dendrite formation in the Li₇La₃Zr₂O₁₂ SE [193]. Results indicated that Li segregation at GBs was energetically driven, manifesting as spontaneous enrichment or depletion governed by local thermodynamic minima. This phenomenon was strongly modulated by temperature and bulk lithium activity. Lower temperatures and elevated lithium concentrations would promote pronounced interfacial segregation. Moreover, horizontally oriented GBs imposed significantly higher kinetic barriers to Li⁺ transport than their vertically aligned counterparts, primarily due to disrupted percolation pathways and increased tortuosity. Critically, GBs containing crack‐like voids acted as preferential nucleation sites for lithium protrusions at the interface, creating localized current hotspots that accelerated dendrite initiation and elevated short‐circuit risk. In this case, controlled thermal annealing above the GB glass transition temperature induced selective amorphization of the boundary regions while preserving the crystalline integrity of the bulk SE lattice (Figure 8f). Although this treatment compromised a reduction in overall ionic conductivity, it effectively suppressed extreme lithium aggregation and eliminated metal protrusion formation at the EEI, thereby improving the interfacial stability.

Despite the achievements in the mechanism elucidation, there are persistent challenges in modeling multiscale ion transport, due to the intricate coupling between static and dynamic disorder. Static distortions arising from cationic/anionic sublattice randomness and lattice strain introduce long‐range heterogeneities that disrupt predictive structure–property mappings, while dynamic fluctuations such as transient octahedral tilting and vacancy migration correlations govern short‐range ion mobility in a nonequilibrium manner. Additionally, the coherence length that reflects the ion transport scale exhibits significant variability across SE classes, particularly in the engineered interfacial nanostructures and mixed glass‐crystal phases. Capturing these complex multiscale transport mechanisms necessitates next‐generation ML simulation frameworks to integrate physical constraints (e.g., thermodynamic factor, defect formation enthalpies), real‐time trajectory data, and adaptive learning algorithms capable of handling nonequilibrium conditions. Besides, the advancements of HT simulations and experimental validation are crucial for the closed‐loop workflow to realize the reliable predictions and rational design strategies for interface engineering.

3.4. Reverse Design of Electrode–Electrolyte Interface

During the battery research, the ultimate effort centers on developing high‐performance LMBs through experimental electrode and electrolyte design. The research conducted by Liang et al. [194] served as a typical case. They formulated a complex quaternary fluorinated ester electrolyte (MA/TFEC/FEC with LiBF₄). This electrolyte was nonflammable and enabled stable cycling under high voltage (>4.3 V) and at −20°C. The performance breakthrough of this study was achieved through the elaborate design of intermolecular solvent–solvent interactions. Nevertheless, this approach relied on sophisticated design logic. While the system fully demonstrated the merits of component synergistic effects, the electrolyte involved multiple solvents, lithium salts and additives, resulting in an extremely vast space for component combinations. Conventional trial‐and‐error testing of each combination individually is impractical, which directly underscores the industry's urgent demand for a ML‐driven inverse design framework. By virtue of the capability of processing high‐dimensional data and uncovering complex intrinsic correlations, ML is emerging as a pivotal tool for revolutionizing electrolyte design. Different from the structure–performance relationships analysis and descriptors extraction that utilize compositional or structural information as input descriptors and the target performances (e.g., cycling stability, ionic conductivity) as the output, while in reverse material design, material performance is designated as the input, with structure and composition serving as the output. In this workflow, experimentally determining the chemical makeup and structure of materials constitutes the pivotal step.

As shown in Figure 9, a rational workflow for the discovery and design of materials should combine the data collection, descriptor extraction, model construction, candidate identification, and experimental validation. Specifically, the workflow is beginning with dataset collection that provides compositional and structural information, and the corresponding output performances. In this process, electrodes and electrolytes with varying compositions and architectures in a defined parameter space are synthesized and tested. Notably, some physical constraints should be considered in the vast parameter space to identify these candidate materials. Meanwhile, theoretical calculations are also the cornerstone to provide the key features (e.g., electronic structures, thermodynamic factors) that affect the output performance for predicting the promising candidates. The obtained experimental and computational results are mutually consistent and reinforcing. Together, they serve as input for ML to elucidate structure–performance relationships and identify functional descriptors. Afterwards, ML models (e.g., regression, neural networks, clustering, classification) trained on the dataset are used to correlate the features and performances, ultimately generating the key descriptors. Crucially, constructing a model to characterize the relationship linking structural features with performance targets is key in this stage. On the basis of the established correlation between descriptors and properties, it is possible to systematically explore and identify novel materials with targeted performance through reverse design. Finally, the predicted materials are synthesized and characterized to validate the prediction. Agreement between prediction and experiment serves as the key criterion for successful material discovery and design. If not, the obtained results are fed back to the database to refine the ML model. The iteration is repeated until the discovery of the optimal candidate. This closed‐loop framework demonstrates high accuracy to provide quantitative structure–performance correlations and provide the fundamental insights into the critical structural factor that governs the interfacial stability.

FIGURE 9.

ML‐based reverse design workflow for EEI in LMBs.

3.4.1. Electrolyte Design

Electrolytes, a vital component of LMBs, exist in two primary forms: liquid and solid. In liquid systems, ion migration is characterized by ongoing solvation or desolvation. Therefore, substantial efforts have focused on the solvation structure regulation, such as advanced electrolyte systems (salts and solvents), perfluorinated electrolytes, and locally concentrated electrolytes. As demonstrated by Chang et al. [82], three key microscopic metrics including solvation coordination number, solvation shell number, and lithium‐ion diffusion coefficient were extracted by the GPR models trained on the 166 formulations composed of 5 lithium salts and 8 ether‐based solvents. Based on the correlations between molecular features and performance, an ideal design space was generated and 11 candidate formulations were proposed (Figure 10a). Further experimental validation demonstrated that these seven candidates exhibited stable performance, which is in high agreement with the predictions. Additionally, two standout formulations were identified including the LiFSI/DEE/LiNFS ternary system, which achieves a CE of 98.32% and the lowest polarization, and the optimized DEE/LiNFS system, which exhibits excellent cycling stability and low polarization characteristics at 1 C. Nevertheless, it should be noted that a fundamental balance exists between solvent‐ and anion‐driven solvation structures in the interfacial solvation environments. Specifically, solvent‐dominated solvation shells, characterized by Li⁺ coordination primarily with solvent molecules and minimal anion involvement, enable rapid Li⁺ transport through the electrolyte bulk. However, they tend to induce the development of an organic‐rich, solvent‐derived SEI, compromising interfacial ion conduction. In contrast, anion‐dominated solvation structures, characterized by anions constituting the primary coordination sphere of Li⁺, are favorable to lower the desolvation energy barrier and favor the formation of inorganic‐rich SEI, thereby improving the reversibility of lithium plating and stripping. Nevertheless, this configuration exhibits inherently low Li⁺ mobility within the electrolyte matrix due to stronger ion‐pairing effects. Traditional electrolyte design relies on specific solvation structures. To break free from this paradigm, Huang et al. [195] proposed an innovative delocalization electrolyte design strategy that creates a more disordered solvation microenvironment, thus reducing the dynamic energy barrier and stabilizing the EEI. In this investigation, a molecular structure database (288 solvents, 99 Li salts) was first systematically compiled by DFT calculations, yielding critical physicochemical descriptors for HOMO, LUMO and solvation energy with Li⁺ (E _solvation). Based on these descriptors, subsequent ML‐based clustering analysis identified 67 target solvent clusters, as well as 19 lithium salts to improve delocalized electrolyte design, both of which exhibited how to balance stability against both oxidation and reduction and moderate interaction with Li⁺. These delocalization configurations exhibited a diverse solvation structure distribution, balancing the respective merits of solvent‐ and anion‐dominated structures. Benefiting from the delocalized solvation chemistry of the electrolyte, the dynamic barriers were mitigated, which promoted Li⁺ diffusion. Additionally, robust and stable interphases were formed. As a consequence, the designed electrolyte endowed the LMB with extended cycling stability and improved energy density of 600 Wh kg⁻¹.

FIGURE 10.

Schematic of multilevel strategies for electrolyte optimization. (a) Identification of the ensemble optimal feature space for electrolyte design across three target properties. Consolidated ensemble optimal feature space constructed by integrating overlapping feature ranges across all three targets. Colored lines represent the eleven experimentally formulated electrolyte designed based on this optimal parameter space. Reproduced with permission [82]. Copyright 2025, Wiley‐VCH GmbH. (b) Li‑Metal Interface Challenges: Liquid vs. Solid‑State Electrolytes. HAML (hybrid AIMD/ML) workflow: geometry optimization and iterative ML‑MD refinement. Reproduced with permission [198]. Copyright 2025, Springer Nature. (c) Workflow of CSP with MTP and AL. Reproduced with permission [199]. Copyright 2025, American Chemical Society. (d) ML‑driven discovery of ionic liquids (High σ, Wide ECW). Reproduced with permission [200]. Copyright 2023, Springer Nature. (e) OTF‐MLMD and AIMD comparative analysis of Li metal/DME‐salt interfacial reactions via energy, charge and bond length data Reproduced with permission [46]. Copyright 2025, Elsevier B.V. and Science Press. (f) Schematic of Li⁺ solvation structure with electrolyte molar ratio quantification. Reproduced with permission [203]. Copyright 2022, American Chemical Society. (g) DiffMix architecture unites GDL and physical laws for mixture property prediction, paired with gradient‐guided robotic electrolyte screening. Reproduced with permission [202]. Copyright 2024, Springer Nature.

With respect to the SSE, the major bottleneck for SSE applications is their limited ionic conductivity. Driven by the need for high ionic conductivity, research efforts have concentrated on the development of multicomponent inorganic SSEs through compositional optimization typically via elemental substitution or doping with base materials. Nevertheless, navigating in the vast parameter space via trial‐and‐error approaches is time and cost prohibitive. Using ML can greatly streamline development cycles and accelerate the discovery of high‐performance SSEs. As demonstrated by Ma et al. [196], a database with 398 garnet solid electrolytes was constructed by combining materials synthesis, structures and performances. Subsequently, a set of ML models were employed to identify critical factors that affect the ionic conductivity, among which Gradient boosting regression exhibited the highest correlation coefficient. Ultimately, the ML model extracted 10 descriptors for materials prediction, and guided the preparation of five garnet‐type solid SSEs. The obtained optimal Li_6.2La₃Zr₂Fe_0.25O₁₂ SSE presented the ionic conductivity of 1.08 × 10⁻⁴ S cm⁻¹ in the experimental testing, which was well consistent with the predicted values of 1.37 × 10⁻⁴ S cm⁻¹, illustrating the accuracy and high efficiency of the ML technology.

Besides the experimental data source, theoretical calculation results can serve as input to train the ML model by aligning with interface reaction mechanisms and atomic‐scale configurations. For instance, MLP was developed for the Li–P–S–Ge–Sn system [197]. This study proposed two strategies including introducing Li vacancies and tailoring the host matrix composition via cation doping, which, based on diffusion barrier calculations and diffusion coefficient analyses, prove crucial for enhancing ionic conductivity. On the basis of the theoretical analysis, Li₁₀Ge(PS₆)₂ was predicted with the highest room‐temperature conductivity. Although the MLP provides critical insights for the mechanisms governing interface reactions, a pivotal problem remains in sophisticated simulations of authentic electrochemical interface reactions with solvation effects and charge transfer taken into account: how to efficiently sample the vast configurational space and ensure sufficient representativeness of MLP training datasets, a factor directly determining the reliability of results. To overcome such limitation, hybrid AIMD combined with MLP was developed to model the Li metal–interfaces and guide the stable interface design [198]. This process is shown in Figure 10b. It was demonstrated that introducing moderate lattice instability was beneficial to accelerate the interface reaction kinetics and thus induce the stable SEI layer formation due to the surface stability anisotropy. Specifically, dopants (e.g., Se, F, and O) in the Li₆PS₅Cl system show great potential for interface stabilization. In addition, inherent crystal structures of SSE also play a fundamental role in ion transport. Kim et al. introduced a phase‐diagram‐guided and MTP‐driven theoretical crystal structure prediction approach to predict novel SSE structures [199]. Figure 10c illustrated the workflow, which consisted of three main stages: generation of initial training sets, iterative structure prediction loop, and final structure prediction. In this framework, the MTP model was iteratively refined through an AL loop, wherein newly identified query configurations selected for maximal information gain were incorporated to align the PES with high‐fidelity DFT reference data. Prediction accuracy is rigorously validated by comparing total energies and radial distribution functions against ab initio benchmarks, ensuring transferability across structural motifs. Building on this diagram, four promising SSE candidates (Li₂SiS₃, Li₂GeS₃, Li₄SiGeS₆, and Li₄SiSnS₆) with novel crystal structures were demonstrated with high ionic conductivity. Mechanistic analysis revealed that these metastable edge‐sharing phases exhibited distinctly enhanced packing efficiency, expanded Li–S₄ sublattice volume, and increased distortion. All of these factors collectively enhance Li⁺ transport, offering crucial guidance for informing the development of advanced SSEs.

Compared with the inorganic SSEs, polymer electrolytes gain much attention due to their excellent interface capability, yet they exhibit certain limitations in practical applications, stemming from insufficient segmental motion kinetics that results in a typical ionic conductivity ranging from 10⁻⁸ to 10⁻⁵ S cm⁻¹. To advance high‐performance polymer electrolytes, as depicted in Figure 10d, Li et al. [200] built a ML framework combined GCNN with quantum calculation to explore ionic liquids for ionic polymer electrolytes. This ML workflow relied on UL to comprehensively investigate the underlying correlations and descriptors in the dataset, followed by SL to filter ionic liquids with desired ionic conductivity. Under the guidance of this workflow, the designed ionic polymer membrane based on LiFSI and C3mpyrFSI enabled the lithium electrode with stable cycling performance without dendrite formation. The fabricated full cell delivered good capacity retention during 350 cycles (>96% at 0.5 C). Overall, the application of ML in electrolyte design has evolved from facilitating interpretable quantitative prediction of performance‐determining rules, to enabling HT screening of novel materials and simulating interfacial phenomena, and ultimately to actively guiding molecular‐level design and device‐level adaptation in complex multi‐component electrolyte systems. This data‐driven rational design paradigm is emerging as the core engine for breaking through the interface bottlenecks of lithium electrodes and developing next‐generation LMBs with high safety and high energy density.

However, the actual performance of electrolytes is governed synergistically by phenomena spanning multiple length scales, including the atomic/molecular, mesoscopic and macroscopic scales. Most traditional ML studies focus solely on molecular descriptors at a single scale, making it difficult to capture cross‐scale synergistic effects. To address this gap, researchers have begun to construct ML‐driven multi‐scale modeling frameworks, integrating theoretical simulations and experimental data, and establish correlations between microscopic structures and the practical performance of batteries. At the atomic/molecular scale, the combination of MLFF and active learning has achieved a milestone breakthrough. Specifically, taking the energy, bond length, force and charge calculated by DFT as training data, researchers learn the high‐dimensional mapping from local atomic environments to potential energy surfaces, and the trained model enables to reach an accuracy close to that of DFT. The uncertainty quantification module of active learning (such as Bayesian estimation) automatically screens key configurations with inaccurate model predictions, which are then calculated via DFT and added to the training set. In this way, the generalization ability of the model is continuously improved with minimal DFT invocation costs. A representative study [46] adopted on‐the‐fly ML molecular dynamics (OTF‐MLMD) with on‐demand active learning. After small‐scale DFT training with merely 340 femtoseconds, the simulation scale was successfully expanded to nearly 300 000 atoms, the largest atomic model ever reported in research on lithium metal batteries. As depicted in Figure 10e, this work realized unified simulations from bulk phases to interfaces at DFT‐level precision, accurately reproduced the reduction mechanism of FSI⁻ anions, revealed the decomposition sequence of LiFSI on lithium metal surfaces (N‐F bonds break first, ultimately generating Li₂O and LiF), and clarified the elemental composition of the SEI: oxygen (50.1%), fluorine (22.2%), sulfur (15.6%) and nitrogen (12.1%). This study trained force fields in real time using small‐scale systems and expanded the simulation cell sizes round by round under the drive of active learning. Eventually, it carried out simulations with millions of atoms at DFT accuracy, achieving an efficient transition from the atomic/molecular scale (DFT‐level energy and force) to the mesoscopic/macroscopic scale (simulations with tens of thousands of atoms or more). When moving up to the mesoscopic scale (solvation structures and interface models), ML plays a vital role in quantitatively correlating electronic and structural features at the molecular scale with the behaviors of solvation clusters at the mesoscopic scale. The electrolyte solvation structure–interface model systematically developed by Ming's research team [201] has laid a solid physical foundation for this purpose. As shown in Figure 10f, the team defined the “Li⁺ solvation structure” as the fundamental unit (Li⁺[solvent]_x[anion]), and introduced the L parameter (interaction strength and configuration between Li⁺ and solvents), B parameter (packing configuration among solvent molecules) and distance parameters to describe interactions between molecules and ions. On this basis, the team adopted techniques such as two‐dimensional nuclear magnetic resonance (2D NMR, including COSY and HOESY) to elucidate weak interactions between solvent molecules as well as between solvents and anions. They also successfully established a quantitative relationship between small‐angle X‐ray scattering (SAXS) results at the mesoscopic scale (cluster sizes ranging from 1 to 100 nanometers) and desolvation energy at the atomic/molecular scale. In short, ML enables the mapping from descriptors at the atomic/molecular scale to geometric parameters of mesoscopic clusters. Meanwhile, it fits the nonlinear relationship between cluster sizes measured by SAXS and desolvation energy (a key property at the atomic/molecular scale). This allows quantitative prediction of mesoscopic cluster behaviors based on atomic and molecular information, or deduction of microscopic interactions reversely from mesoscopic measurement results, building a bridge between the microscopic and mesoscopic scales. At the macroscopic scale, ML efficiently navigates the vast chemical space through active learning and Bayesian optimization to directly optimize the macroscopic performance of batteries. As depicted in Figure 10g, Zhu et al. [202] developed a differentiable geometric deep learning framework named DiffMix in the study. This framework inputted molecular species, component concentrations and environmental conditions into GNNs, and learned the coefficients of the Redlich‐Kister thermodynamic mixing theory and the Vogel‐Fulcher‐Tammann transport equation, so as to predict macroscopic properties such as ionic conductivity in an end‐to‐end manner. Guided by simulation trajectories from this framework, the ionic conductivity of the electrolyte was enriched by more than 18.8% after only 10 rounds of robotic experiments on the Clio platform. The differentiability of the framework provides gradient information to guide the adjustment of formulas along the steepest ascent direction, forming a closed‐loop cross‐scale design workflow that ranges from molecular structures to macroscopic performance (ionic conductivity) and further to automated experimental verification. In consequence, a set of preliminary cross‐scale research methodologies based on ML have been constructed in the field of electrolyte design. In the future, researchers may further embed physical information into multi‐scale models and develop multimodal contrastive learning architectures, thereby realizing the seamless integration of atomic‐scale calculations (DFT), mesoscopic‐scale simulations (MD) and macroscopic experimental tests, and ultimately accomplishing a fundamental paradigm shift: away from single‐descriptor‐dependent predictive modeling to reverse design across multiple scales.

3.4.2. Electrode Design

Apart from the electrolyte design, surface protection, and microstructural design of electrodes are also key directions for stabilizing lithium interface. At the atomic level, synthesis conditions, lattice parameters, atomic arrangement, and coordination environments collectively dictate ion diffusion and interfacial reactions, thereby exerting a profound influence on battery behavior. Leveraging its advantages in handling high‐dimensional data, ML can uncover hidden structure–performance relationships and mechanisms to inform the systematic design of advanced electrode architectures. As shown in Figure 11a, calculations of lithium diffusion rate and adsorption energy via MLIP combined with DFT computational method provided an initial assessment guidance for interface design. It was demonstrated that copper was favorable to suppress dendrite formation, while lithium‐indium alloy tends to cause the dendrite generation [204]. Besides, ML provides an advanced methodology to transfer the materials screening for intelligent design of interface layers. Zhang et al. [1] developed a framework that combines ML with HT quantum‐mechanical simulations to guide the design of a synthetic inorganic–organic hybrid SEI layer on the Li‐metal electrode to mitigate dendrite growth and enhance cycling stability. As shown in Figure 11b, (1) 2400 organic molecules were first collected from the PubChem database and 83 linear‐chain, F‐containing organic molecules that favored self‐assembling monolayers were screened out. (2) gradient boosted decision tree subsequently mapped the correlations between structural and electronic descriptors with Li⁺ diffusion to evaluate the organic molecules. (3) 128 candidates exhibiting robust stability and reduced Li⁺ diffusion barriers were predicted, among which 8 candidates with desired structural characteristics were identified. Moreover, the systematical analysis illustrated that the terminal group of –CH₃ in synergy with the head units (–OF, –CF, –NF, –SF, and –PF) contributed to the low diffusion barrier, which provided critical insights to guide the design and discovery of promising SEI layers for practical use in LMBs. Additionally, for the complex high‐entropy alloys interphase layer with chemical disorder or atomic‐scale structural disorder, ML‐driven DFT calculations revealed the important role of Zn in lithiophilicity, and Al in lithiophilicity and mechanical stability, as well as Fe, Co, and Ni in providing magnetism [205]. Ultimately, the optimal atomic ratio of Al, Zn, Fe, Co, and Ni to be 7:5:4:3:1 for designing high‐entropy alloys SEI was determined. As validated by the further experiments, symmetric batteries based on this interlayer exhibited stable cycling for 2400 h at 1 mA·cm⁻² and 1 mAh·cm⁻².

FIGURE 11.

Multiscale simulation and ML‐driven electrode design schematics. (a) Diffusion barriers obtained using NEB method with MTP (blue lines) and DFT (green lines). The first two are Li (100), and the latter two are Cu (111) and LiIn (100), respectively. Reproduced with permission [204]. Copyright 2024, AIP Publishing. (b) Relationship between Li⁺ BE and HEA element proportions. Reproduced with permission [1]. Copyright 2023, Elsevier Inc. (c, d) Workflow for virtual search space construction and closed‑loop AL for optimizing electrolytes in anode‑free batteries. Reproduced with permission [152]. Copyright 2025, Springer Nature. (e) QRNN versus OPLS₄ MD: Li⁺ RDFs and cluster distribution statistics of LiPF₆/EC electrolyte. Reproduced with permission [210]. Copyright 2025, Royal Society of Chemistry. (f) Da_II‐SOC curves of control and HOLE anodes. Reproduced with permission [211]. Copyright 2023, Elsevier B.V. (g) Multiscale Hierarchical Electrode Design Framework Based on Network Layers and Interaction Networks. Reproduced with permission [212]. Copyright 2021, Elsevier B.V. All rights reserved. Nano‐micro‐macro scale: Reproduced with permission [213]. Copyright 2015, Springer Nature. Reproduced with permission [214]. Copyright 2019, Wiley‐VCH. Model parameters: Reproduced with permission [215]. Copyright 2020, Elsevier B.V. Simulation outcomes: Reproduced with permission [216]. Copyright 2016, Elsevier B.V. All rights reserved. Material synthesis: Reproduced with permission [217]. Copyright 2017, Wiley‐VCH. Multiscale model: Reproduced with permission [218]. Copyright 2018, The Electrochemical Society.

Although the lithium metal electrode offers high energy density compared to LIBs, the inherent drawbacks of lithium metal, such as high chemical reactivity and poor reversibility of deposition/stripping, severely limit the cycle life of batteries. As a result, existing researches and practical use commonly implemented excess lithium metal to extend cycle life of LMBs, which compromises energy density and poses additional safety risks. Recently, so‐called anode‐free configuration featuring only an unmodified current collector on the cathode side attracted much attention to maximize the energy density. Nevertheless, since no metallic lithium source is present to offset capacity loss [206], it necessitates the electrolyte design to induce SEI formation for enabling long‐term cycling stability of anode‐free LMBs. In this scenario, the role of ML has been further elevated: shifting from directly designing interlayers to indirectly constructing and optimizing interfaces by designing electrolyte components. Especially, the robustness of Bayesian methods with noisy data stems from their native handling of uncertainty estimation [207]. This characteristic allows them to identify higher‐success‐potential hypotheses despite experimental variability interference, and guides researchers to explore the underexplored experimental directions. For the anode‐free LMB optimization, a sequential Bayesian experimental design was employed to pinpoint top candidates under conditions of both data scarcity and label noise [152]. As shown in Figure 11c,d, in the framework, a vast chemical space composed of unannotated small organic molecules from e‐Molecules and PubChem databases was utilized to define the electrolyte optimization space. Further filtration was performed to refine the search space by removing chemical moieties and low ionic conductivity. Subsequently, AL based on a Bayesian model was leveraged to identify the optimal candidate. For the unlabeled electrolytes recommended by the AL framework, experimental tests were conducted and iteratively fed back the obtained battery cycling data to the model for performance optimization. After six rounds of iterative experiments, four high‐performance ether‐based electrolyte formulations were efficiently identified from the vast chemical space, confirming the validity of ML in assisting novel electrode design.

From this perspective, ML in electrode design has evolved from assisting in the rapid evaluation of fundamental material properties, to actively and precisely optimizing the multi‐dimensional parameters of interlayers, and finally to extending to addressing data‐scarce scenarios through AL. This capability lays a solid methodological foundation for the rational design of future advanced battery electrodes with high energy density, high safety, and long cycle life. One core challenge persists: accurately simulating realistic battery operating conditions remains difficult with current technology, even with recent progress.

This issue is particularly prominent in anode‐free batteries, whose performance degradation is especially sensitive to interfacial instability. Notably, anode‐free batteries do not contain excess metallic lithium for capacity compensation. Their cyclic capacity fading is mainly governed by multiple failure mechanisms, including active lithium loss, continuous growth of the SEI film and elevated interfacial impedance. Relying solely on traditional ML to screen electrolyte formulas with high stability makes it difficult to accurately identify the dominant failure causes in different systems. With the continuous iteration of ML and time‐series analytics, such data‐driven tools have expanded their scope in anode‐free battery research from simple performance screening to in‐depth research scenarios such as accurate identification of interfacial failure causes and degradation modes. By integrating interpretable models like SHAP temporal attribution and long short‐term memory (LSTM) networks, researchers can quantitatively analyze the influence weight of each characteristic parameter across the full battery lifespan, so as to achieve precise traceability of battery failure behaviors. Piao et al. [208] built ML models leveraging professional battery expertise. Drawing on tens of thousands of cycling tests and tens of millions of battery data points, they clarified the correlation between the early electrochemical behaviors of lithium anodes and their subsequent failure. Using merely less than 2% of the data from the first two cycles, the model predicted three categories of battery failure, categorized as kinetic degradation, reversibility degradation, and synergistic degradation. Feature analysis of the model verified that lithium nucleation and surface layer thickening are the core determinants of failure modes. This strategy can be extended to sodium anodes, zinc anodes and even anode‐free batteries, offering critical support for tracing the failure mechanisms of batteries. Meanwhile, Liu et al. [209] conducted elaborate testing and analysis across atomic, electrode and full‐cell scales on high‐energy‐density 450 Wh kg⁻¹ Ah‐level anode‐free pouch cells, revealing two typical failure modes: the “sudden death” mode (primarily caused by brittle and rigid SEI films that readily form truffle‐like dead lithium, directly triggering cell failure) and the “gradual decay” mode (originating from flexible SEI films with mushroom‐shaped active lithium growing alongside them, leading to continuous capacity fading of the battery). The study further discovered that for high‐energy anode‐free batteries, the elastic deformability of SEI films served as a more reliable metric for evaluating the performance stability of SEI in high‐energy‐density anode‐free cells compared with the conventional Young's modulus. The above technical system improves the full‐chain application of ML in anode‐free batteries, and provides solid theoretical support and technical guidance for the long‐term stable operation and practical engineering translation of anode‐free batteries.

Beyond time‐series failure analysis, the fundamental insights of electrode interfacial stability also span multiple scales. The in‐depth integration of MLFF and active learning technologies has first unlocked pioneering advances at the atomic/molecular scale. Li et al. [210] constructed an AL simulation framework based on the MTP/MACE MLFF adaptable to crystalline‐amorphous mixed SEI, as illustrated in Figure 11e. The training dataset was iteratively refined via amorphous sampling and extrapolation criteria to balance simulation accuracy and computational cost, and the framework was extended to atomic simulations of various SEI components such as Li₂CO₃ and Li₂EDC. By comparing and measuring the diffusion difficulty of lithium ions through multiple atomic simulation methods, it was found that intact inorganic layers severely hinder lithium transport, while structural defects could significantly accelerate Li ion migration. This work observes a special lithium relay conduction pathway at the lithium metal–SEI interface at the atomic scale, and quantitatively interprets the regulation law of temperature on the coordination structure of SEI based on radial distribution functions at different temperatures. Capable of simulating the microscopic evolution of SEI over long durations with high precision, this framework bridges the correlation between microscopic atomic structures and macroscopic battery performances. At the mesoscopic scale, as shown in Figure 11f, Goel et al. [211] employed particle swarm optimization (PSO) to automatically calibrate the parameters of the Newman porous electrode model. They also introduced the second Damköhler number Da_II (characteristic diffusion time /characteristic reaction time) as a universal metric to optimize the channel spacing of laser‐patterned holey electrodes. The electrodes delivered the optimal performance with a channel width of 55 µm when Da_II ≈ 1 under 4 C charging. Meanwhile, a semi‐analytical method was built to lower the 3D simulation time to 180 s, realizing the rapid mapping from mesoscopic structure (channel spacing) to macroscopic performance (charging capacity). Alternatively, at the macroscopic system level, the CHAIN–MED framework (Figure 11g) proposed by Gao et al. [212] further integrated multi‐scale models (ranging from nanometer, micrometer to macroscopic scales to manufacturing scales, or covering molecular dynamics, 3D reconstruction, P2D models and digital twins) with multi‐source databases. Algorithms including random forests, neural networks and Bayesian optimization were utilized to mine structure–activity relationships, transforming electrode design from empirical trial‐and‐error into a closed‐loop hybrid optimization mode driven by the combination of data and physics. The above studies collectively demonstrate that ML has gone beyond descriptor‐performance fitting at a single scale, and formed a cross‐scale technical chain ranging from the microscale (MLFF for reaction mechanism elucidation), mesoscopic scale (PSO combined with Da_II for electrode geometry optimization) to the macroscopic scale (closed‐loop prediction and optimization). It provides a scalable paradigm to bridge longstanding gaps obstructing multi‐scale research in electrode design.

4. Conclusion and Outlook

LMBs with their ultrahigh theoretical capacity and low redox potential, represent the highly promising rechargeable batteries to achieve unprecedented energy densities for next‐generation energy storage systems [219]. However, their practical industrialization is greatly hindered by uncontrolled Li dendrite growth and highly reactive interfaces. The advent of data‐driven ML methodologies has significantly revolutionized interface stabilization in the development of LMBs. This review establishes a full‐process systematic framework integrating ML for interface engineering, balancing experimental investigations and computational simulations, and discussing the existing achievements and unresolved challenges in the industry. It sequentially constructs an end‐to‐end closed‐loop simulation‐experiment workflow that connects HT experiments, DFT calculations, molecular dynamics simulations and ML modeling. The paper comprehensively analyzes key thermodynamic and kinetic factors governing electrode design, electrolyte optimization and interfacial stability regulation, delivering an all‐round interpretation of the complete EEI system. Moreover, it screens feature descriptors and combines physical laws to conduct targeted inverse design of new materials without blind trial‐and‐error tests. Furthermore, this review achieves cross‐scale integration spanning atomic, mesoscopic and macroscopic scales, systematically elaborating how ML enables in‐depth mechanistic insights into critical electrochemical processes, including dendrite evolution, SEI film formation and ion transport behavior. Finally, it prospectively proposes a development path combining generative AI with automated experimental platforms, which is expected to form a complete closed loop linking material prediction, synthesis and characterization. Collectively, these innovations create a brand‐new research paradigm, deepen the fundamental understanding of ML‐assisted interface engineering, and offer valuable references for follow‐up research. Despite the impressive advances, two fundamental bottlenecks persist: imperfect data and limited mechanism understanding (Figure 12). To overcome such limitations, following feasible potential solutions deserve much attention, including data standardization and integration, multiscale modeling combination, multimodal model development, and intelligent systems establishment that combine AI, computations, and experiments.

FIGURE 12.

Challenges and future directions for ML integration research framework for lithium metal battery interfaces engineering.

1. Establishing a robust database as ML input. ML‐driven battery design relies on the establishment of an integrated database capable of fusing high‐dimensional, multi‐source, heterogeneous and sparse data. In this regard, the recent conceptual framework proposed by Cai et al. [220] is highly enlightening. One core viewpoint is that establishing supporting standards and a standardized evaluation system is a prerequisite for constructing a data‐driven research paradigm. Furthermore, they emphasize that converting compositional, structural and performance indicators into machine‐readable formats (standardized binary information represented by “0” and “1”) constitutes a critical step to guarantee data interoperability and the generalization capability of models. Most existing data are sourced from literature mining, HT experiments and theoretical calculations, while individual data source is subject to its inherent limitations. Literature‐based text mining requires extensive natural language processing to extract structured information from unstructured texts. However, electrode formulations, electrolyte compositions and performance testing conditions (e.g., battery capacity, current density, cycling protocol) are typically scattered and non‐standardized, leading to the compromised model accuracy and biased results. Moreover, most structured materials databases predominantly focus on basic physicochemical descriptors (e.g., crystallographic features), rarely containing critical electrochemical performance metrics, compelling researchers to rely on indirect proxies that compromise predictive fidelity. Although HT experimental workflows are expected to produce huge and consistent datasets, their significant demands on time and experimental resources inevitably constrain the attainable dataset volume. Additionally, computational methods such as DFT calculations limit to simulating interface reactions in a realistic environment. To overcome these data obstacles, it is highly demanded to establish a comprehensive data library based on standardized formats to enable cross‐laboratory interoperability and collaborative sharing to contribute proprietary data. For example, an industrywide consortium could create open‐access databases in which manufacturers and researchers contribute process and performance data under unified feature definitions. Proprietary information could be protected through secure measures, such as federated learning [221]. Notably, besides the positive outcomes, negative or mediocre results also deserve much attention as inputs fed to ML models, helping to fill critical data gaps and ensure the full diversity.

2. Coupling multiscale modeling to elucidate the in‐depth mechanism. The stability of Li EEI is governed by hierarchical interplay factors spanning atomic (e.g., chemical bonds and electronic structure), nanoscopic (e.g., defects and phase interfaces), mesoscopic scale (e.g., multiphase complexity), and macroscopic scale (e.g., cell‐level impedance). Nevertheless, current ML models are typically trained on isolated scales, while lacking a universal predictive framework that connects multi‐scales phenomena, and struggling to fully and effectively elucidate the cross‐scale structure–property relationships. Besides, most ML‐driven methodologies employ static structural parameters, failing to obtain the impacts of dynamic information about reaction states (e.g., SEI intermediates, local strain fields, harmonic vibrations, solvation structure transformation) on real‐time Li⁺ transport and subsequent Li plating and stripping. This limitation will be further exacerbated by the utilization of equilibrium‐phase simulations (e.g., DFT) that neglect the nonequilibrium, time‐dependent behavior intrinsic to battery operation. Bridging this gap necessitates multi‐scaling modeling involving high‐quality training data obtained from precise synthesis and characterization techniques (e.g., in situ spectroscopy or atomic‐resolution imaging), and sophisticated simulation methods like DFT and AIMD simulations. For instance, MD simulations yield critical atomic‐scale descriptors, including Li⁺ diffusion coefficients and solvation configurations. These outputs inform phase‐field models that simulate void nucleation, grain boundary segregation, and dendrite morphology evolution under local stress and ion flux gradients. Subsequently, the resultant phase‐field simulation data are incorporated into finite element models to quantify cell‐level electrochemical responses such as impedance rise and short‐circuit risks. Leveraging these cross‐scale simulation insights, ML models are developed to establish quantitative mappings from mesoscopic structural patterns to macroscopic device‐level performance metrics. Most importantly, iterative experimental characterization work continuously refines model parameters and validates simulation predictions, forming a closed and self‐consistent modeling‐experiment feedback loop. Such cross‐scale integration from the atomic upward to nanoscale, mesoscopic, and macroscopic simulations is critical to constructing a coherent computational framework, laying the foundation for reinforcing structure–property relationships and advancing the design and optimization of highly stable interface.

3. Developing multi‐modal models for inverse deign of stable interface. To reconcile the persistent discrepancy between experimental observations and computational predictions, a multi‐modal contrastive learning architecture is beneficial to provide the structured and meaningful parameters for advancing the interface engineering in LMBs. An example can be demonstrated through the synergistic integration of (1) extracting atomic configurations (e.g., Li coordinates, band structure of electrolyte) by DFT, (2) observing the dynamic reaction (e.g., SEI formation, Li nucleation and deposition) via in situ characterizations, and (3) implementing convolutional neural networks and deep learning algorithms to analyze the multi‐scale structural evolution of electrode materials. A coherent, multi‐scale picture of battery degradation ranging from atomic ion transport and mesoscale nucleation to 3D network and interfacial evolution is emerging. This collective understanding provides a robust theoretical basis for the design of advanced LIBs. For instance, combined graph convolutional neural networks (GCNNs) with quantum calculations, polymer electrolytes with suitable ionic liquids were efficiently found from more than 2200 candidates [200]. Such a framework effectively aligns experimental data and simulation results, improving the transferability of mechanism insights to real‐world conditions. This facilitates the model to derive systematic mapping relations from thermodynamic and kinetic viewpoints, significantly promoting the fundamental understanding of interfacial phenomena. By uncovering latent relationships across data modalities, ML offers novel perspectives for interface engineering. Unified embeddings, for instance, can reveal SEI compositions linked to high ionic conductivity across both simulated and experimental datasets. Additionally, specific microstructural deposits' morphology observed in tomography images corresponds to computational predictions of the dendrite formation trends. This iterative, ML‐guided loop integrated experiments and theoretical calculations is expected to provide a new perspective for the rational design of high‐performance Li electrode–electrolyte interfaces.

4. Improving the interpretability and reliability of ML models. Limited interpretability and generalizability of ML models constitute a pivotal bottleneck hindering their practical deployment in LMBs interface engineering. To date, most ML‐driven approaches rely on black‐box algorithms to yield critical insights into the underlying reaction mechanisms, rendering it a challenge to assess prediction confidence and arduous to trust or directly guide the design directions. Directly implementing entirely transparent modeling based on explicit rules or physical principles allows for a clear delineation of the rational process behind predictions [41]. Nevertheless, it is limited to explain the dynamic reaction and metastable interphases. Going further, it is favorable to incorporate domain knowledge or physical constraints into ML algorithms to improve interpretability and prediction accuracy. For instance, Rhyu et al. [206] employed a systematic feature design framework for battery life prediction. Based on domain knowledge, early‐stage life predictions with an average error below 10% were achieved, showing the potential of feature constraints on improving the accuracy of models. Moreover, overcoming the trust barrier and limited generalizability is also worth much attention. In this case, physics‐informed or knowledge‐integrated hybrid models can be embedded into domain‐specific constraints (e.g., reaction mechanisms, conservation laws) to improve both physical consistency and predictive reliability. Besides, XAI tools, such as SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model‐agnostic Explanations), and sensitivity analysis, can help researchers better understand the effects of input features on model outputs, thereby strengthening trust in model judgments [222]. A representative study based on a deep learning‐based framework, along with the SHAP analysis to analyze the feature importance for detecting lithium‐plating‐type defects in LIBs effectively enhanced the model's interpretability [223]. Results align with the electrochemical mechanisms of Li plating, achieving an overall accuracy of 97.42%. Ultimately, we emphasize that to truly move beyond the “black box optimization” paradigm, future ML‐based research on battery interfaces must integrate embedded physical constraints and ensure the interpretability of model outputs. Once further combined with LLMs and automated laboratories, a complete intelligent research and development (R&D) system can be established. This criterion becomes even more critical when explainable models are deployed in the intelligent closed‐loop R&D system described below.

5. Building an intelligent platform integrating AI, experiments and computations. AI methods provide powerful tools for data processing and data mining to automatically extract structure–performance relationships. In the future, ML techniques are expected to evolve to autonomously guide research design, analyze and identify the valuable insights from experiments and theoretical calculations. To realize these capabilities, they should be integrated into a synergistic platform that combines ML algorithms, simulation tools, materials synthesis, characterization, performance testing, and real‐time feedback, enabling iterative optimization and efficient design of advanced electrodes and electrolytes with minimal human intervention. Recent advances in AI‐empowered closed‐loop experimental platforms have demonstrated that the integration of ML prediction, robotic experimentation, and autonomous optimization algorithms enables HT electrolyte screening, battery assembly, and electrochemical testing within several hours. Such efficient workflows drastically shorten the experimental cycle, which conventionally requires weeks or even months via traditional methods. Unlike conventional ML which merely relies on previous literature and static data for analysis, emerging automated intelligent platforms boast unique strengths. For example, ALBATROSS fully automates battery preparation and testing, delivering more stable experimental data than manual operation [224]. BRINE can autonomously adjust electrolyte formulas repeatedly while recording all data from successful and failed experiments, addressing the common problems in the battery industry of insufficient data and the practice of only keeping records of successful trials [225]. Auto‐MISCHBARES, a millimeter‐scale HT platform, adopts asynchronous network scheduling to run multiple sets of experiments in staggered parallel. It automatically controls experiment quality, analyzes data, and stores all raw experimental data in a standardized manner for easy subsequent retrieval and traceability [226]. Even so, AI‐driven automated battery research still has shortcomings at present. Inconsistent cross‐platform standards lead to data silos; insufficient prior experimental knowledge restricts the convergence effect of active learning; and the accuracy of online characterization and analysis for microscopic side reactions at interfaces is inadequate. Therefore, standardizing the battery testing conditions is of high significance for the subsequent optimization of intelligent platforms. As a typical example, LLMs, reinforcement learning and autonomous laboratories can be integrated into an intelligent system, in which LLMs systematically extract and organize multi‐dimensional information, RL algorithms learn optimal parameter region according to the outcome feedback, while autonomous laboratories can automatically perform subsequent experiments [227, 228, 229]. This closed‐loop workflow systematically screens vast parameter space, directs automated synthesis and characterization workflows, and imperatively refines design hypotheses. This self‐driving platform enables the rapid convergence of candidate materials, extraction of powerful descriptors from generated database. Moreover, such a system overcomes the challenges of method standardization and mechanism distortion, effectively bridging the gap between experimental results and theoretical calculations, yielding the in‐depth understanding of thermodynamics and kinetics. Consequently, the transition from conceptual design to prototype verification is significantly accelerated, promoting the realization of high‐performance batteries that transcend the limitations of conventional empirical approaches.

6. Generative AI for inverse design of battery interfaces. Besides the aforementioned intelligent R&D frameworks, generative AI models have emerged as powerful tools for accelerating the research and development of battery interface engineering. Generative AI refers to a class of AI technologies capable of learning the underlying distribution patterns of data and generating novel samples. Its core strength lies in designing previously unknown molecules and material structures from scratch, forming a closed‐loop pipeline that integrates iterative molecular generation, performance evaluation, and structural optimization. The mainstream approaches include GANs, VAEs and diffusion models, each with distinct features. GAN generates target molecules in a targeted manner relying on attribute labels; VAE maps molecules to latent spaces and regulates molecular performance through interpolation and parameter optimization [39]. Diffusion models can bias generation toward desired targets, making them highly relevant to the optimization of battery materials [48]. In the EEI design of lithium batteries, generative AI accelerates rational design at multiple levels. It can independently design additives capable of forming stable SEI films, formulate polymer electrolytes and artificial interfacial layers, develop new solvents that meet electronic performance requirements, construct electrode microstructures to evaluate interfacial properties, and realize “inverse design” for specific interfaces when combined with closed‐loop automated platforms. For instance, MatterGen, a diffusion model for inorganic crystalline materials developed by the Zeni team, employed a customized diffusion process tailored to the characteristics of crystalline materials. Pre‐trained on a dataset of over 600 000 crystals, it leveraged adapters and classifier‐free guidance technology to design materials targeted by elemental composition, crystal symmetry, physical properties, and other criteria. The proportion of high‐quality new structures it produced was more than twice that of previous state‐of‐the‐art models, with a substantial reduction in structural energy deviation. Moreover, in specific ternary systems, it identified four new structures using merely over 10 000 samples, boasting far higher efficiency than traditional random search methods [230]. Concurrently, Song and colleagues proposed Crystal Synthesis Large Language Models (CSLLM). Supported by three fine‐tuned large models, it performed crystal synthesizability assessment, synthesis method classification, and precursor recommendation in sequence. The team built a dataset containing tens of thousands of positive and negative samples and adopted concise material strings to represent crystal information. After fine‐tuning, the model achieved a synthesizability prediction accuracy of 98.6%, outperforming traditional thermodynamic and kinetic screening methods in both accuracy and computing speed. To facilitate practical applications, this framework has been developed into a visual tool that enables one‐click completion of multiple synthesis‐related analyses and physical property predictions for crystals [231]. Collectively, generative AI can fundamentally overcome the inherent limitations of traditional trial‐and‐error experimentation by efficiently exploring the vast, untapped chemical and structural space of battery materials. It expedites the R&D and innovation of new materials and supporting systems related to lithium battery interfaces across various aspects, ranging from small molecules and polymer materials to electrode microstructures and automated iterative optimization.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.