Skip to main content
NCBI home page
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2026 Jul 9;65(37):e6825619. doi: 10.1002/anie.6825619

A Quantitative Electrostatic Potential Descriptor Enables Deep Learning‐Accelerated Discovery of High‐Performance Lithium‐Ion Battery Electrolytes

Kun Han 1, Yu Lou 1, Junfeng Li 2, Chaocang Weng 3, Chenglong Wang 1,✉, Wenjie Mai 4, Jinliang Li 4,✉, Guang Yang 1, Likun Pan 1,✉
PMCID: PMC13549032  PMID: 42424155

ABSTRACT

Rational electrolyte design for high‐energy‐density lithium‐ion batteries (LIBs) urgently demands precise and quantitative molecular descriptors of solvation power to enable deep learning (DL)‐accelerated screening, yet such descriptors remain lacking. Here, we introduce the electrostatic potential ratio |ESPmin|/ESPmax (ESPratio) as a quantitative descriptor capturing the balance between electron‐donating and electron‐accepting capacities, and identify a solvation modulation zone (0.9 < ESPratio < 2.4) through unsupervised clustering of 344 molecules encompassing 196 experimentally reported LIB electrolyte molecules. By combining this descriptor with self‐supervised pre‐trained DL models fine‐tuned on small experimental datasets, we enable hierarchical screening of ∼106 PubChem molecules and prioritize electrolyte candidates from previously unexplored chemical space. Experimental evaluation of representative candidates, including TBDN and PIV as co‐solvents and additional nitrile‐containing molecules as electrolyte additives, confirms that the ESPratio‐guided workflow can enrich chemically meaningful electrolyte candidates for high‐voltage Li||LiCoO2.

Keywords: electrolyte design, electrostatic potential, lithium‐ion batteries, machine learning, solvation structure

A quantitative solvation modulation zone (0.9 < |ESPmin|/ESPmax < 2.4) is established through data‐driven clustering of 344 molecules. Machine learning‐accelerated hierarchical screening of ∼106 candidates enables experimental validation of five screened molecules in high‐voltage Li||LiCoO2 cells, with tert‐butylmalononitrile (TBDN) emerging as a previously unexplored electrolyte co‐solvent that achieves 99.4% Coulombic efficiency and 80.9% capacity retention over 150 cycles at 4.6 V.

graphic file with name ANIE-65-e6825619-g004.webp

1. Introduction

Increasing the energy density of lithium‐ion batteries (LIBs) is imperative for the rapid evolution of electric vehicles [1, 2, 3]. Among widely used cathode materials, LiCoO2 (LCO) is particularly prominent owing to its superior theoretical capacity (274 mAh g−1) and high operating voltage [4]. However, commercial electrolytes such as LP30 exhibit limited oxidation stability (∼4.3 V) and suffer from severe side reactions at the high‐voltage LCO cathode interface [5, 6, 7]. Consequently, optimizing electrolyte design represents a core strategy for enhancing the performance of LIBs [8, 9, 10]. As the primary component of electrolytes, solvents fundamentally dictate electrolyte properties through their solvating power and electrochemical stability [11, 12, 13]. Therefore, solvent screening emerges as a central approach for rational electrolyte design. However, the development of high‐performance electrolytes currently relies heavily on inefficient trial‐and‐error experiments [14, 15, 16]. To address this challenge, machine learning (ML) has emerged as a transformative approach to accelerate the screening process of electrolytes [17, 18, 19, 20, 21, 22, 23]. Recent studies further show that AI‐driven approaches are increasingly used not only for materials and electrolyte screening, but also for extracting mechanistic insight from complex battery datasets [24, 25].

The overall electrochemical performance of an electrolyte is fundamentally governed by the solvation structure of lithium ions (Li+), including ionic conductivity, interfacial stability, and electrode/electrolyte compatibility [26, 27, 28, 29]. This structure directly affects Li+ transport kinetics, desolvation energy barrier, as well as the composition and stability of the cathode–electrolyte interphase (CEI) [30, 31]. Essentially, the solvation structure results from the interaction strength between solvent molecules and Li+. Consequently, developing a reliable metric to quantitatively evaluate the solvents’ solvation ability is a crucial theoretical basis for understanding the structure–property relationships in electrolytes and rationally designing high‐performance electrolyte systems. Currently, the donor number (DN) is widely regarded as an indicator of a solvent's electron‐donating ability in the electrolyte field [20, 32, 33], defined as the enthalpy change when solvent molecules coordinate with the standard Lewis acid SbCl5 in dilute solution. However, the sole use of DN to characterize solvent–Li+ interactions in LIB systems has inherent limitations. DN is defined by the interaction between a solvent and SbCl5, a large Lewis acid probe whose steric and electronic properties differ substantially from those of Li+; this mismatch means that DN fails to fully capture the actual Li+ solvation sheath structure and desolvation kinetics. Therefore, complementing DN with additional molecular descriptors that more directly reflect the electronic features of solvent–Li+ interactions is therefore essential for precise, high‐throughput electrolyte screening.

The electrostatic potential (ESP), derived from quantum chemical calculations, provides a direct link between molecular electronic structure and intermolecular interactions. Negative ESP regions generally correspond to electron‐rich sites that can coordinate Li+, whereas positive ESP regions reflect electron‐deficient regions that influence salt dissociation and desolvation‐related interactions. Recent electrolyte studies have established the relevance of ESP extrema for solvent and additive design. For example, the ESPmin‐ESPmax map has been used to qualitatively distinguish antisolvents, weakly solvating solvents, and strongly solvating solvents [34], while Cai et al. recently included ESPmin/ESPmax‐derived quantities as molecular input features in an ML model for aqueous zinc‐ion battery additives [35]. ESPmin and ESPmax have also been used together with interaction energy to evaluate Li+ solvation behavior in ionic‐liquid electrolytes [36]. These independent studies support the view that ESP extrema and their derived quantities encode chemically meaningful information for electrolyte design. However, existing ESP‐based approaches remain limited for automated high‐throughput LIB solvent screening. In previous studies, ESPmin and ESPmax were mainly used as qualitative map coordinates, independent molecular descriptors, or input features in supervised performance models. Such uses confirm the value of ESP analysis, but they do not provide a single dimensionless decision variable with defined numerical boundaries. In particular, visual classification on an ESPmin‐ESPmax plot cannot be directly applied as a reproducible filter for million‐scale molecular libraries. A quantitative descriptor that converts ESP extrema into an actionable screening criterion is therefore still needed.

In this work, we convert the physical insight contained in ESPmin and ESPmax into a quantitative screening descriptor by defining ESPratio = |ESPmin|/ESPmax. This dimensionless ratio describes the relative balance between electron donating and electron accepting electrostatic characteristics, rather than their absolute difference. Through data‐driven clustering of experimentally reported LIB electrolyte solvents, we identify a solvation modulation zone (SMZ), defined as 0.9 < ESPratio < 2.4, which serves as a programmable criterion for high throughput molecular screening. The overall strategy is summarized in Figure 1, starting from the limitations of conventional DN descriptors and proceeding through ESPratio definition, ML‐accelerated hierarchical screening of approximately 106 PubChem molecules, and experimental validation in high‐voltage Li||LCO cells. This screening framework identifies several electrolyte candidates from previously unexplored chemical space. Subsequent concentration‐dependent electrochemical evaluation covered both co‐solvent and additive‐type formulation tests, with tert‐butylmalononitrile (TBDN) and pivalonitrile (PIV) examined as co‐solvents, and octanedinitrile (ODN), trimethylsilyl cyanide (TMSCN), and 2‐methylglutaronitrile (2‐MPDN) examined as electrolyte additives.

FIGURE 1.

FIGURE 1

Open in a new tab

Overall workflow for quantitative ESP‐based solvent screening and experimental validation. (a) Schematic comparison of the DN probe and Li+ illustrating the size difference between the DN probe molecule and Li+, motivating the development of complementary ESP‐based descriptor for LIB applications. (b) Definition of the ESPratio (|ESPmin|/ESPmax) and the data‐driven solvation modulation zone (0.9 < ESPratio < 2.4) derived from unsupervised clustering of five molecular clusters (C0–C4), with ethylene carbonate (0.913) and 2‐methyltetrahydrofuran (2.313) serving as lower and upper bounds, respectively. (c) Three‐stage hierarchical screening of ∼106 PubChem molecules using ML models (Uni‐Mol) for oxidation potential, DN, ESPmin, and ESPmax, progressively narrowing candidates to 353 molecules. (d) Experimental validation of representative screened candidates in Li||LCO cells. For visual clarity, two co‐solvent candidates, tert‐butylmalononitrile (TBDN) and pivalonitrile (PIV), are shown in the schematic; in total, five screened molecules were experimentally evaluated, including two co‐solvents and three electrolyte additives.

2. Results and Discussion

2.1. Dataset Construction and Statistical Analysis

To establish a quantitative framework for solvent screening, we constructed three datasets encompassing diverse molecular properties essential for electrolyte performance evaluation. The oxidation potential dataset comprises 93 264 molecules retrieved from the Materials Project (MP) database [37]. The DN dataset incorporates 195 molecules compiled from the R. Stenutz DN database. The ESP dataset contains 344 molecules for which ESP parameters were computed via density functional theory (DFT) calculations. Among these, 196 are reported LIB electrolyte molecules systematically compiled from the literature via Web of Science, covering nearly a decade of published research. Two independent contemporaneous studies reported similar‐scale compilations [18, 21], corroborating that our dataset achieves near‐complete coverage of the experimentally explored chemical space to date. Within this subset, molecules with well‐established electrochemical performance were further annotated as validated high‐performance solvents or antisolvents based on documented experimental evidence (detailed procedures for dataset construction and curation are provided in the Supporting Information Section 1). Initial statistical analysis of the three datasets reveals distinct compositional characteristics and substantial overlap in molecular properties across chemical categories (Figure 2a–c). The oxidation potential dataset from the MP exhibits predominant representation of aromatic compounds. In contrast, the DN and ESP datasets encompass broader functional diversity, including carbonates, ethers, sulfones, nitriles, and esters that are directly relevant to LIB electrolytes. Notably, both ESPmin and ESPmax distributions show overlapping ranges among different functional groups, indicating that univariate thresholds on individual parameters cannot effectively discriminate solvent performance across diverse chemical classes.

FIGURE 2.

FIGURE 2

Open in a new tab

Dataset characterization and ESP distribution analysis. (a–c) Histograms of oxidation potential, DN, and ESPmax across chemical categories in the three constructed datasets. (d) ESP maps of representative high‐performance solvents (PC, DMC, DEC, and MP) with corresponding ESPmin and ESPmax values. (e) Box plots comparing the |ESPmin| and ESPmax distributions across the full dataset. (f) 2D scatter plot of |ESPmin| versus ESPmax for experimentally validated solvents, grouped by chemical class.

To probe the electronic origins of solvation capability, we examined the ESP distributions of representative molecules with demonstrated high performance in LIB electrolytes (Figure 2d). The molecular surfaces reveal a consistent electronic signature wherein electron‐rich regions (negative ESP) substantially dominate over electron‐deficient regions (positive ESP). This electron‐donating predominance, manifested as |ESPmin| > ESPmax, characterizes the Lewis base functionality essential for coordinating Li+ through electrostatic attraction to negative potential regions. Statistical analysis of the full dataset shows that |ESPmin| values span from 0.8 to 3.6 eV, while ESPmax ranges from 0.4 to 2.8 eV (Figure 2e), demonstrating the broad chemical diversity. The comparison across these structurally diverse molecules suggests that effective solvents share this qualitative electronic asymmetry regardless of their specific chemical structure, indicating that the relative balance between electron‐donating and electron‐accepting capabilities is critical for effective solvation.

To assess whether this qualitative insight could be translated into quantitative screening criteria, we mapped the experimentally validated high‐performance solvents in the two‐dimensional (2D) |ESPmin| versus ESPmax space (Figure 2f). Most critically, these molecules with proven efficacy in LIB electrolytes do not cluster within a well‐defined region. Instead, they scatter across the ESP landscape, occupying territories that span nearly the entire observed range. This scattered distribution indicates that no simple boundary in the 2D ESP space can effectively encompass all validated solvents while excluding poor performers. Notably, different chemical categories exhibit tendencies toward distinct zones, with carbonates generally clustering separately from ethers, which in turn segregate from sulfones. This category‐specific spatial segregation reveals that the relationship between ESP parameters and solvation performance is primarily modulated by chemical class. These results demonstrate that the governing physics of solvation cannot be captured by simple thresholds on individual ESP parameters or their direct combinations in 2D space. The pronounced chemical class dependencies and scattered distribution of high‐performance molecules suggest that a normalized descriptor capturing the relative electrostatic balance holds the key to predicting solvation capability across diverse molecular structures.

2.2. ESP Ratio Descriptor via Data‐Driven Clustering

The scattered distribution of high‐performance solvents in absolute ESP space motivates the use of a normalized descriptor that captures solvation‐relevant electrostatic balance while reducing chemical‐class dependence. We define ESPratio as |ESPmin|/ESPmax, where |ESPmin| reflects electron‐donating character associated with Li+ coordination and ESPmax reflects electron‐accepting electrostatic character associated with salt dissociation and desolvation‐related interactions. ESPratio is therefore used here as a solvation‐relevant screening descriptor rather than as a direct predictor of cell‐level performance metrics. A ratio was selected because it describes the relative balance between the two ESP extrema in a dimensionless form, whereas a difference descriptor remains tied to the absolute ESP scale. A detailed comparison between ESPratio and |ESPmin|− ESPmax is provided in Supporting Information Section 2.4 and Figure S1.

To further evaluate the descriptor, we analyzed both its statistical association with known solvent behavior and its physicochemical connection to Li+ binding. Across 42 labeled molecules, ESPratio shows a strong association with the binary solvent or antisolvent label, giving a point‐biserial correlation coefficient of r = 0.750 and a Spearman correlation coefficient of ρ = 0.692 (Figure S2). Li+ binding energy calculations further show that most good solvents within the SMZ exhibit appreciable Li+ affinity, whereas antisolvents outside the SMZ generally show weaker binding (Figure S3a). The aromatic false positives toluene and m‐xylene show negligible Li+ binding despite falling nominally inside the SMZ, and they are removed by the upstream DN filter owing to their low predicted DN values (Figure S3b). These results indicate that ESPratio captures solvation‐relevant electrostatic balance and functions most appropriately as a candidate‐enrichment descriptor within a hierarchical screening workflow. Because ESPratio is a scalar descriptor, it does not explicitly encode all steric effects or the spatial separation between ESPmin‐ and ESPmax‐localized regions. Molecules selected by ESPratio should therefore be regarded as candidates for further verification rather than as guaranteed solvation modulators. In the present workflow, newly screened candidates are further assessed through additional ESP calculations, electrochemical testing, and molecular or structural characterization.

To systematically identify regions of chemical space associated with effective solvation behavior, we employed unsupervised clustering analysis of the 344‐molecule ESP dataset. The molecules were first encoded as Morgan fingerprints and embedded into a 2D representation using t‐distributed stochastic neighbor embedding (t‐SNE) (detailed parameters are provided in Supporting Information Section 2) [14, 38, 39]. K‐means clustering was subsequently applied to partition the embedded space, with the optimal number of clusters determined through silhouette coefficient analysis (Figure S4) [40]. This workflow was used to stratify the ESP dataset into local molecular‐fingerprint regions rather than to assign molecules into strict functional‐group classes. As shown in Figure 3a, the five clusters show clear but non‐exclusive compositional trends. Cluster C0 is enriched in ester and carbonate molecules, C2 is a mixed solvent‐relevant region containing carbonates, sulfones, other molecular skeletons, and compact cyclic ethers such as 2‐MeTHF, whereas C4 is enriched in ether‐type molecules. A quantitative summary of the molecular‐skeleton composition of C0, C2, and C4 is provided in Table S1. These regions cover major structural families of established LIB solvents and contain the experimentally established boundary molecules used in SMZ construction. In contrast, C3 contains both labeled good solvents and antisolvents, indicating that cluster membership alone is insufficient for solvent classification. The labeled‐set distribution further supports this conservative domain selection (Figure S5). C0, C2, and C4 contain 21 of the 33 labeled good solvents and only one of the nine antisolvents, corresponding to a selected‐domain label purity of 95.5%. In contrast, C1 and C3 contain eight of the nine antisolvents. Although C3 also contains labeled good solvents, its higher antisolvent burden makes it less suitable as a conservative structural domain for high‐throughput candidate enrichment. Therefore, C0, C2, and C4 are used here as conservative solvent‐relevant structural domains for subsequent ESPratio analysis, rather than as deterministic labels of solvent performance.

FIGURE 3.

FIGURE 3

Open in a new tab

t‐SNE visualization of the ESP dataset with k‐means clustering and ESPratio mapping. (a) Five‐cluster partition of 344 molecules in Morgan‐fingerprint‐based t‐SNE space. The clusters represent local structural‐similarity regions rather than strict functional‐group classes. (b) ESPratio mapped onto the same embedded space using a blue‐to‐red color scale, with ethylene carbonate (0.913) and 2‐MeTHF (2.313) identified as the lower and upper anchors of the Solvation Modulation Zone.

Within these solvent‐relevant structural domains, we computed ESPratio for each molecule and visualized this descriptor on the clustered space using a color scale where blue indicates low ratios and red denotes high ratios (Figure 3b). The experimentally validated solvents within these domains span a bounded ESPratio range rather than concentrating at a single value. Ethylene carbonate provides the lower anchor with Although the DN prediction ESPratio = 0.913, whereas 2‐methyltetrahydrofuran provides the upper anchor with ESPratio = 2.313. With slight extension to accommodate numerical precision and boundary‐adjacent cases, we define the solvation modulation zone as 0.9 < ESPratio < 2.4. Thus, the final criterion combines solvent‐relevant structural‐domain selection with a continuous ESPratio window. This zone should be interpreted as a data‐driven empirical criterion for enriching LIB organic electrolyte solvent candidates, rather than as a universal boundary for all electrolyte systems. Its numerical limits are derived from the currently available experimental LIB solvent space and may be refined as additional validated solvent data become available.

2.3. ML‐Accelerated Property Prediction and Hierarchical Screening Strategy

Following the establishment of the quantitative ESPratio criterion, we developed deep learning (DL) models to enable rapid property prediction across extensive chemical libraries. The Uni‐Mol architecture was selected as the foundation for all predictive models owing to its capacity to encode both 2D molecular connectivity and three‐dimensional conformational information, which are essential for capturing the geometric and electronic features that govern ESP distributions and redox behavior [41, 42]. Four distinct regression models were trained to predict oxidation potential, DN, ESPmin, and ESPmax using the datasets described in the preceding sections (detailed fine‐tuning procedures for these four models are provided in Supporting Information Section 3 and Table S2). For the ESP descriptors, a joint multi‐target training approach was initially explored but ultimately abandoned in favor of independent single‐target models, as separate training yielded lower mean absolute errors (MAE) for both endpoints (see Supporting Information Section 3.4). As demonstrated by the parity plots in Figure 4, the oxidation potential model attained MAE of 0.07 eV on the training set and 0.08 eV on the test set. The ESPmin and ESPmax models similarly demonstrated high precision, with test set MAE of 0.09 and 0.10 eV, respectively, thereby confirming their reliability for high‐throughput screening applications. The DN model exhibited a larger test error of 3.59 kcal mol−1 relative to the training error of 1.86 kcal mol−1. Although the DN prediction error remains higher than those of the OP and ESP models, it is still competitive with recent DN prediction models trained on similarly limited datasets, despite their use of extensive molecular feature engineering (Figure S6) [20, 32, 33]. To further benchmark the pretrained‐model strategy, we added lightweight Random Forest, Gradient Boosting, and XGBoost baselines using Morgan fingerprints, basic RDKit descriptors, and their concatenation. These comparisons are summarized in Supporting Information Section 3.6, Figure S7, and Tables S3,S4, and support the advantage of fine‐tuning pretrained molecular representations for electrolyte‐property prediction under limited labeled data.

FIGURE 4.

FIGURE 4

Open in a new tab

Parity plots of the four Uni‐Mol regression models for (a) oxidation potential, (b) DN, (c) ESPmin, and (d) ESPmax, with training (orange) and test (gray) sets shown separately and mean absolute errors annotated in each panel.

Prediction uncertainty was further considered when interpreting the screened candidate pool. Following the donor‐number‐based classification used by Ma et al., [20] the DN > 10 kcal mol−1 threshold was adopted as a coarse lower‐bound filter to exclude molecules with clearly insufficient solvating ability, rather than as a high‐precision classifier for molecules located close to the cutoff. For the ESP‐based criterion, an error‐propagation analysis is provided in Supporting Information Section 3.5, showing that ESPratio uncertainty mainly affects molecules located near the SMZ boundaries. Accordingly, the OP, DN, and ESPratio criteria are used here as candidate‐enrichment filters rather than deterministic predictors of final electrochemical performance.

Upon validation of the predictive models, we implemented a three‐stage hierarchical screening protocol to identify promising solvent candidates from ∼106 small molecules in the PubChem database (detailed procedures for database acquisition and data cleaning are detailed in Supporting Information Section 1). The first stage applied an oxidation potential threshold of 4.0 V versus lithium metal to select molecules with high‐voltage stability. Because the DL models predict properties of isolated molecules in the gas phase, and molecular oxidation potentials in actual electrolyte environments typically exceed their intrinsic values, the 4.0 V criterion provides a conservative margin to ensure electrochemical stability under practical operating conditions. This filtering step reduced the initial library to 4982 molecules, representing less than one percent of the original population. The second stage screened for DN values exceeding 10 kcal mol−1 to maintain adequate solvating power for lithium salt dissociation, further narrowing the candidate pool to 1468 molecules [43]. The third stage applied the quantitative ESPratio criterion, retaining only molecules with 0.9< ESPratio < 2.4 to identify candidates within the solvation modulation zone. This final filter yielded 353 molecules that simultaneously satisfy all three requirements for high‐voltage electrolyte applications. The five‐cluster partition of the 344‐molecule ESP dataset in t‐SNE space (Figure 5a) reveals that the final candidates predominantly distribute within the selected solvent‐relevant structural domains C0, C2, and C4, consistent with the structural‐domain gate used in SMZ construction. As illustrated in Figure 5b, the transition from DN screening to ESP ratio filtering represents a critical refinement step. Although the 1468 molecules passing the DN threshold all possess sufficient solvating power according to the conventional descriptor, the ESPratio descriptor eliminated approximately 76 percent of these candidates, reducing the pool to 353 molecules. This reduction is not used to claim that ESPratio is superior to DN. Rather, it shows that ESPratio contributes complementary electronic information, and the final candidate set is defined by the intersection of oxidation potential, DN, and ESPratio criteria. Because this final set is defined by set intersection, it is independent of the order in which the individual filters are applied.

FIGURE 5.

FIGURE 5

Open in a new tab

Hierarchical screening of the PubChem library and chemical space visualization of the final candidates. (a) t‐SNE map of the 344‐molecule ESP dataset colored by cluster identity. (b) ESPratio distribution within the selected solvent‐relevant structural domains C0, C2, and C4, with open circles marking experimentally validated solvents. (c) t‐SNE visualization of ∼106 PubChem molecules colored by chemical class, with the 353 final candidates highlighted as red stars.

2D t‐SNE visualization of the complete million‐molecule library, colored by chemical class, reveals the compositional diversity of the screening process (Figure 5c). The original PubChem database exhibits broad representation across functional groups, including esters (n = 188,477), ethers (n = 243,496), nitriles (n = 73,888), carbonates (n = 5,432), sulfones (n = 25,575), and phosphates (n = 3,209). Notably, the final 353 candidates demonstrate pronounced enrichment of nitrile‐containing molecules, which constitute a substantially larger fraction than their representation in the initial library. This compositional shift reflects the convergence of high oxidation stability inherent to the electron‐withdrawing cyano group and moderate DN values that fall within the effective range for Li+ solvation. Representative molecules among the 353 final candidates exhibit diverse molecular architectures, including linear nitriles, branched nitriles with tertiary carbon centers, and halogenated variants. The unexpected prominence of nitriles, which have received relatively limited attention compared to conventional carbonate and ether electrolytes, underscores the discovery potential of unbiased computational screening guided by quantitative electronic descriptors.

Among the 353 molecules satisfying all screening criteria, 49 were commercially available. Literature verification showed that 36 of these commercially available molecules had not been previously reported as electrolyte components in electrochemical energy‐storage systems. The complete commercially available candidate list is provided in Figure S8. To further audit the screening output, we recalculated DFT‐level ESP descriptors for selected commercially available molecules recovered by the workflow that were found to have prior electrolyte‐related reports. Most of these audited reported molecules remain within the SMZ, whereas methyl hydrogen carbonate and 1,3,5‐pentanetricarbonitrile fall below the pre‐defined lower boundary after DFT recalculation (Table S5 and Figure S9). These cases indicate that the SMZ should be used as a candidate‐enrichment criterion rather than as a perfect classifier, and that post‐screening candidates require downstream verification rather than post hoc boundary recalibration. To select molecules for experimental validation from this commercially accessible pool, we considered both chemical‐space representativeness and practical experimental feasibility. The molecular similarity principle, widely used in chemoinformatics and molecular property prediction, states that structurally similar molecules are more likely to exhibit similar physicochemical or biological properties [44, 45]. Accordingly, we visualized the commercially available candidates together with previously studied electrolyte solvents in Morgan‐fingerprint‐based t‐SNE space and used local proximity as a qualitative guide for experimental follow‐up (Figure 6a). Candidate prioritization considered chemical‐space representativeness, novelty status, structural diversity, commercial availability, reagent cost, and handling feasibility, without assigning formulation roles at the screening stage. Figure 6 highlights five representative molecules from the broader experimental follow‐up set, and Figure 6b shows the molecular structures of the two candidates emphasized in the main electrochemical discussion. Their practical formulation roles and electrochemical outcomes were resolved only after concentration‐dependent testing. Additional experimentally examined candidates, including less effective cases, are documented in the Supporting Information and in the released candidate dataset. DFT‐level ESP maps of the five experimentally evaluated molecules are provided in Figure S10 to visualize the electrostatic features of the co‐solvent and additive candidates selected from the screening workflow.

FIGURE 6.

FIGURE 6

Open in a new tab

Molecular similarity analysis and representative experimental follow‐up candidates. (a) t‐SNE map of the 49 commercially available candidates alongside previously studied electrolyte solvents. Five representative molecules from the experimental follow‐up set are highlighted. (b) Molecular structures, SMILES strings, and CAS numbers of representative highlighted candidates. Their practical use as co‐solvents or electrolyte additives was determined subsequently through concentration‐dependent electrochemical testing. TBDN, tert‐butylmalononitrile. PIV, pivalonitrile. ODN, octanedinitrile. TMSCN, trimethylsilyl cyanide. 2‐MPDN, 2‐methylglutaronitrile.

2.4. Experimental Validation in High‐Voltage Li||LCO Cells

To experimentally assess the screened candidates under high‐voltage battery conditions, TBDN was first evaluated in Li||LCO cells using LP30 as the baseline electrolyte. The LP30+TBDN formulation was tested between 3.0 and 4.6 V at 0.5 C (1C = 270 mAh g−1) without additional cathode modification or cell‐level optimization beyond electrolyte formulation. As shown in Figure 7a, cells using the baseline LP30 electrolyte exhibited rapid capacity decay, with an average Coulombic efficiency (CE) of 98.9% and a capacity retention of 57.0% within 100 cycles. In contrast, cells using LP30+TBDN maintained an average CE of 99.4% and a capacity retention of 80.9% over 150 cycles. These results indicate that the screened molecule can improve the high‐voltage compatibility of the LP30‐based electrolyte under the present testing conditions. NMR measurements were performed to examine whether TBDN changes the electrolyte environment while remaining compatible with LiPF6. The 19F NMR spectra of LP30 and LP30+TBDN show the characteristic PF6− resonances without additional impurity peaks, indicating that TBDN does not induce obvious LiPF6 decomposition under the examined conditions (Figure 7b). The slight chemical‐shift variation in the 19F spectra suggests a modified local environment around PF6−. Consistently, the 7Li NMR spectra show a corresponding shift after TBDN addition (Figure 7c), indicating that the Li+ coordination environment is altered in the LP30+TBDN electrolyte [46]. Molecular dynamics simulations were then used to further probe the solvation structure. In the baseline LP30 electrolyte, Li+ is mainly coordinated by carbonate oxygen atoms and PF6− fluorine atoms (Figure 7d–f). After introducing TBDN, the nitrile group participates in the Li+ solvation environment, while carbonate oxygen and PF6− fluorine coordination are retained (Figure 7e–g) [47]. The RDF‐derived coordination numbers show that TBDN contributes to the primary Li+ coordination environment rather than simply acting as an inert diluent [48]. This solvation‐structure modulation is consistent with the NMR results and provides a molecular‐level basis for the improved electrochemical behavior. The cycled LCO cathodes were further characterized after 30 cycles at 4.6 V. SEM images show that the LCO particles cycled in LP30+TBDN retain a smoother and more intact surface morphology, whereas the particles cycled in LP30 show more pronounced surface damage and cracking (Figure 7 h,i). XRD patterns further indicate better preservation of the layered LCO structure in the TBDN‐containing electrolyte (Figure 7j). The I(003)/I(104) ratio of the LCO cathode cycled in LP30+TBDN is 2.33, higher than that of the cathode cycled in LP30 (1.64), which is consistent with better retention of the layered LCO structure during high‐voltage cycling [49]. XPS analysis was used to examine the cathode electrolyte interphase formed after cycling. The survey and high‐resolution spectra show that the CEI contains fluorinated inorganic species and carbon‐containing interphase components after cycling in LP30 and LP30+TBDN (Figure 7k–m) [50]. The change in the relative interphase composition after TBDN addition indicates that the screened molecule affects interfacial chemistry at the high‐voltage LCO surface. Taken together with the cycling, NMR, MD, SEM, and XRD results, these data show that TBDN improves the LP30‐based electrolyte by modulating Li+ solvation and stabilizing the cathode interface under 4.6 V operation.

FIGURE 7.

FIGURE 7

Open in a new tab

Characterization of Li||LCO cells using LP30 and LP30+TBDN electrolytes. (a) Cycling performance of Li||LCO cells using LP30 and LP30+TBDN at 0.5 C (1C = 270 mAh g−1) in the voltage range of 3.0–4.6 V. (b) 19F NMR and (c) 7Li NMR spectra of LP30 and LP30+TBDN electrolytes. MD simulation snapshots of the solvation structures in (d) LP30 and (e) LP30+TBDN. Radial distribution functions and coordination numbers in (f) LP30 and (g) LP30+TBDN. SEM images of LCO electrodes after 30 cycles using (h) LP30+TBDN and (i) LP30. (j) XRD patterns and (k–m) XPS spectra of LCO electrodes after 30 cycles using LP30 and LP30+TBDN.

To examine whether the screening workflow could identify additional experimentally responsive molecules, other screened candidates were evaluated under LP30‐based formulation conditions at 4.5 V. The corresponding cycling, spectroscopic, simulation, morphological, structural, and interfacial data for PIV are provided in Figures S11–S21. Additional cycling tests for ODN, TMSCN, and 2‐MPDN are shown in Figures S22–S24. Four less effective candidates, Bromoacetonitrile (BAN), 2‐Bromopropionitrile (2‐BPN), 3‐Bromopropionitrile (3‐BPN), and 3‐Cyanopropyltrichlorosilane (3‐CPTCS), are shown in Figures S25–S28. These results, together with Table S6, document both improved and unsuccessful experimental outcomes and support the use of the screening workflow as a candidate‐enrichment strategy rather than a deterministic predictor of final cell performance. The complete screening output, including predicted molecular properties, candidate annotations, commercial availability, and all molecules examined in our laboratory regardless of electrochemical outcome, will be released through the GitHub repository to support independent validation and community‐driven exploration of AI‐suggested electrolyte candidates.

3. Conclusion

This work establishes a new descriptor for ESP‐based solvent screening by defining the solvation modulation zone as 0.9< ESPratio < 2.4 through unsupervised clustering analysis of 344 molecules, including 196 experimentally reported LIB electrolyte solvents. The agreement between computational screening and experimental validation supports the use of ESPratio as a candidate‐enrichment descriptor within the present hierarchical workflow. Rather than establishing a one‐to‐one predictor of cell‐level performance, the framework links molecular electrostatic balance to solvation‐relevant properties and provides a practical route for prioritizing electrolyte candidates. It should be noted that the SMZ boundaries established in this work are data‐driven empirical criteria derived from and validated within the domain of organic liquid electrolyte molecules for LIB applications. Extension to other electrochemical systems, such as sodium‐ion batteries, aqueous electrolytes, solid‐state electrolytes, or structurally distinct molecular frameworks, would require re‐derivation of the zone boundaries from analogous experimental datasets. Nevertheless, the descriptor definition and screening workflow provide a transferable strategy for building system‐specific solvation criteria.

Author Contributions

Kun Han: conceptualization, methodology, writing – original draft, data curation. Yu Lou: data curation, validation. Junfeng Li: investigation. Chaocang Weng: investigation. Chenglong Wang: formal analysis, software, writing – review and editing. Wenjie Mai: investigation. Jinliang Li: investigation, formal analysis, writing – review and editing. Guang Yang: software. Likun Pan: methodology, writing – review and editing, supervision, conceptualization.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Detailed information on dataset construction, DFT and molecular dynamics calculations, molecular representation, t‐SNE and clustering analysis, ESPratio validation, Uni‐Mol fine‐tuning, model evaluation, baseline comparisons, error‐propagation analysis, candidate annotation, experimental follow‐up status, and supplementary electrochemical and materials characterization data is provided in the Supporting Information. The authors have cited additional references within the Supporting Information [20, 32, 51, 52].

Supporting File: anie73524‐sup‐0001‐SuppMat.docx.

ANIE-65-e6825619-s001.docx (5.3MB, docx)

Acknowledgements

The authors are thankful for the support from Grand NeoBay Development Fund (202509‐KW‐1.1.1‐001) and Guangxi Science and Technology Program (2024AB08156).

Contributor Information

Chenglong Wang, Email: clwang@phy.ecnu.edu.cn.

Jinliang Li, lijinliang@email.jnu.edu.cn.

Likun Pan, Email: lkpan@phy.ecnu.edu.cn.

Data Availability Statement

The data and code that support the findings of this study are available at https://github.com/Hank3824/li_electrolyte_molecule_screen and https://github.com/Hank3824/Reading-notes-code/tree/main/paper%202. The first repository contains the cleaned code, curated datasets, screening outputs, candidate annotations, commercially available candidate list, the 36 previously unexplored candidates, and experimental‐status tables. The second repository contains the Materials Project and PubChem API data‐retrieval scripts used to construct the upstream training and screening datasets. Additional information is available from the corresponding author upon reasonable request.

References

  • 1. Xu G. S., Jiang M. X., Li J. L., et al., “Machine Learning‐accelerated Discovery and Design of Electrode Materials and Electrolytes for Lithium Ion Batteries,” Energy Storage Mater 72 (2024): 103710, 10.1016/j.ensm.2024.103710. [DOI] [Google Scholar]
  • 2. Cheng X. B., Liu H., Yuan H., et al., “A Perspective on Sustainable Energy Materials for Lithium Batteries,” Susmat 1 (2021): 38–50, 10.1002/sus2.4. [DOI] [Google Scholar]
  • 3. Celadon A., Sun H. H., Sun S. H., and Zhang G. X., “Batteries for Electric Vehicles: Technical Advancements, Environmental Challenges, and Market Perspectives,” Susmat 4 (2024): e234. [Google Scholar]
  • 4. Manthiram A. and Goodenough J. B., “Layered Lithium Cobalt Oxide Cathodes,” Nat Energy 6 (2021): 323–323, 10.1038/s41560-020-00764-8. [DOI] [Google Scholar]
  • 5. Zhang F. C., Qin N., Li Y. Z., et al., “Phytate Lithium as a Multifunctional Additive Stabilizes LiCoO2 to 4.6 V,” Energy & Environmental Science 16 (2023): 4345–4355, 10.1039/D3EE01209C. [DOI] [Google Scholar]
  • 6. Cheng F. Y., Xu J., Wei P., et al., “Interface Engineering via Regulating Electrolyte for High‐Voltage Layered Oxide Cathodes‐Based Li‐Ion Batteries,” Advancement of Science 10 (2023): 2206714, 10.1002/advs.202206714. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Li X. L., Lin Z. F., Jin N., et al., “Perovskite‐Type SrVO3 as High‐Performance Anode Materials for Lithium‐Ion Batteries,” Advanced Materials 34 (2021): 2107262, 10.1002/adma.202107262. [DOI] [PubMed] [Google Scholar]
  • 8. Fan X. L. and Wang C. S., “High‐voltage Liquid Electrolytes for Li Batteries: Progress and Perspectives,” Chemical Society Reviews 50 (2021): 10486–10566, 10.1039/D1CS00450F. [DOI] [PubMed] [Google Scholar]
  • 9. Martins V. L., “Advances on Liquid Electrolytes for Li‐ion and Li Metal Batteries,” Current Opinion in Electrochemistry 38 (2023): 101241, 10.1016/j.coelec.2023.101241. [DOI] [Google Scholar]
  • 10. Jie Y. L., Ren X. D., Cao R. G., Cai W. B., and Jiao S. H., “Advanced Liquid Electrolytes for Rechargeable Li Metal Batteries,” Advanced Functional Materials 30 (2020): 1910777, 10.1002/adfm.201910777. [DOI] [Google Scholar]
  • 11. Zheng Z., Zhang X. Q., Zhan D. H., et al., “Elucidating the Solvating Power of Solvents for Designing High‐Performance Electrolytes in Lithium Batteries,” ACS Energy Letters 10 (2025): 4962–4982, 10.1021/acsenergylett.5c02291. [DOI] [Google Scholar]
  • 12. Yamada Y., Wang J. H., Ko S., Watanabe E., and Yamada A., “Advances and Issues in Developing Salt‐concentrated Battery Electrolytes,” Nat Energy 4 (2019): 269–280, 10.1038/s41560-019-0336-z. [DOI] [Google Scholar]
  • 13. Yamada Y., Furukawa K., Sodeyama K., et al., “Unusual Stability of Acetonitrile‐Based Superconcentrated Electrolytes for Fast‐Charging Lithium‐Ion Batteries,” Journal of the American Chemical Society 136 (2014): 5039–5046, 10.1021/ja412807w. [DOI] [PubMed] [Google Scholar]
  • 14. Gao Y. C., Yao N., Chen X., Yu L. G., Zhang R., and Zhang Q., “Data‐Driven Insight into the Reductive Stability of Ion–Solvent Complexes in Lithium Battery Electrolytes,” Journal of the American Chemical Society 145 (2023): 23764–23770, 10.1021/jacs.3c08346. [DOI] [PubMed] [Google Scholar]
  • 15. Sendek A. D., Cubuk E. D., Antoniuk E. R., Cheon G., Cui Y., and Reed E. J., “Machine Learning‐Assisted Discovery of Solid Li‐Ion Conducting Materials,” Chemistry of Materials 31 (2019): 342–352, 10.1021/acs.chemmater.8b03272. [DOI] [Google Scholar]
  • 16. Chen J. W., Zhang D. M., Zhu L., et al., “Hybridizing Carbonate and Ether at Molecular Scales for High‐energy and High‐safety Lithium Metal Batteries,” Nature Communications 15 (2024): 3217, 10.1038/s41467-024-47448-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Zheng Z., Liu X. Y., Zhang X. Q., et al., “Deciphering Coulombic Efficiency of Lithium Metal Anodes by Screening Electrolyte Properties,” Angewandte Chemie International Edition 64 (2025): e202507387, 10.1002/anie.202507387. [DOI] [PubMed] [Google Scholar]
  • 18. Gao Y. C., Yuan Y. H., Huang S. Z., et al., “A Knowledge–Data Dual‐Driven Framework for Predicting the Molecular Properties of Rechargeable Battery Electrolytes,” Angewandte Chemie International Edition 64 (2025): e202416506, 10.1002/anie.202416506. [DOI] [PubMed] [Google Scholar]
  • 19. Chen X., Liu M. K., Yin S. Q., Gao Y. C., Yao N., and Zhang Q., “Uni‐Electrolyte: An Artificial Intelligence Platform for Designing Electrolyte Molecules for Rechargeable Batteries,” Angewandte Chemie International Edition 64 (2025): e202503105, 10.1002/anie.202503105. [DOI] [PubMed] [Google Scholar]
  • 20. Jia Q. Q., Liu H. G., Wang X. P., et al., “Machine Learning‐Driven Mass Discovery and High‐Throughput Screening of Fluoroether‐Based Electrolytes for High‐Stability Lithium Metal Batteries,” Angewandte Chemie International Edition 64 (2025): e202424493, 10.1002/anie.202424493. [DOI] [PubMed] [Google Scholar]
  • 21. Kumar R., Vu M. C., Ma P. Y., and Amanchukwu C. V., “Electrolytomics: A Unified Big Data Approach for Electrolyte Design and Discovery,” Chemistry of Materials 37 (2025): 2720–2734, 10.1021/acs.chemmater.4c03196. [DOI] [Google Scholar]
  • 22. Ma P. Y., Kumar R., Wang K. H., and Amanchukwu C. V., “Active Learning Accelerates Electrolyte Solvent Screening for Anode‐free Lithium Metal Batteries,” Nature Communications 16 (2025): 8396, 10.1038/s41467-025-63303-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Xu G. S., Li Y., Li J. F., et al., “Toward Stable Zinc Anode: An AI‐Assisted High‐Throughput Screening of Electrolyte Additives for Aqueous Zinc‐Ion Battery,” Angewandte Chemie International Edition 64 (2025): e202511389. [DOI] [PubMed] [Google Scholar]
  • 24. Han Z. Y., Zhou J. Q., Lu G. X., et al., “Uncovering Battery Electrochemical Mechanisms by Artificial Intelligence,” National Science Review 12 (2025): nwaf442, 10.1093/nsr/nwaf442. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25. Lu G. X., Han Z. Y., Shi L., et al., “Decoding Single‐crystal Lithium Growth through Solid Electrolyte Interphase Omics,” Nature Communications 16 (2025): 9323, 10.1038/s41467-025-62166-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26. Cui B. C. and Xu J. J., “Enabling Rational Electrolyte Design for Lithium Batteries through Precise Descriptors: Progress and Future Perspectives,” Journal of Materials Chemistry A 13 (2025): 8223–8245, 10.1039/D4TA07449A. [DOI] [Google Scholar]
  • 27. Xiao P. T., Yun X. R., Chen Y. F., et al., “Insights into the Solvation Chemistry in Liquid Electrolytes for Lithium‐based Rechargeable Batteries,” Chemical Society Reviews 52 (2023): 5255–5316, 10.1039/D3CS00151B. [DOI] [PubMed] [Google Scholar]
  • 28. Qin T., Yang H. Y., Wang L., et al., “Molecule Design for Non‐Aqueous Wide‐Temperature Electrolytes via the Intelligentized Screening Method,” Angewandte Chemie International Edition 63 (2024): e202408902, 10.1002/anie.202408902. [DOI] [PubMed] [Google Scholar]
  • 29. Husch T. and Korth M., “Charting the Known Chemical Space for Non‐aqueous Lithium–air Battery Electrolyte Solvents,” Physical Chemistry Chemical Physics 17 (2015): 22596–22603, 10.1039/C5CP02937F. [DOI] [PubMed] [Google Scholar]
  • 30. Rao Y. X., Li X., Zhao S. Q., et al., “Fluorinated Electrolyte Formulations Design Enabling High‐voltage and Long‐life Lithium Metal Batteries,” Nano Energy 123 (2024): 109362, 10.1016/j.nanoen.2024.109362. [DOI] [Google Scholar]
  • 31. Dai T., Kouoi X., Reynaud M., et al., “Antiperovskite Active Materials for Metal‐ion Batteries: Expected Advantages, Limitations, and Perspectives,” Energy Storage Mater 68 (2024): 103363, 10.1016/j.ensm.2024.103363. [DOI] [Google Scholar]
  • 32. Gao Y. Q., Qi S. G., Li M. R., et al., “Machine Learning Assisted Prediction of Donor Numbers: Guiding Rational Fabrication of Polymer Electrolytes for Lithium‐Ion Batteries,” Angewandte Chemie International Edition 64 (2025): e202411437, 10.1002/anie.202411437. [DOI] [PubMed] [Google Scholar]
  • 33. Luo H. R., Gou Q. Z., Zheng Y. J., et al., “Machine Learning‐Assisted High‐Donor‐Number Electrolyte Additive Screening toward Construction of Dendrite‐Free Aqueous Zinc‐Ion Batteries,” ACS Nano 19 (2025): 2427–2443, 10.1021/acsnano.4c13312. [DOI] [PubMed] [Google Scholar]
  • 34. Wu Y. Z., Hu Q., Liang H. M., et al., “Electrostatic Potential as Solvent Descriptor to Enable Rational Electrolyte Design for Lithium Batteries,” Advanced Energy Materials 13 (2023): 2300259, 10.1002/aenm.202300259. [DOI] [Google Scholar]
  • 35. Cai B., Sun Y. X., Yu F., et al., “Machine‐Learning‐Assisted Screening of Electrolyte Additives for Aqueous Zinc‐Ion Batteries via a Molecular Descriptor Framework,” Angewandte Chemie International Edition 65 (2026): e21382, 10.1002/anie.202521382. [DOI] [PubMed] [Google Scholar]
  • 36. Liu J., Zhang S. Z., Yao Y. C., et al., “Electrostatic Potential Characteristics of Ionic Liquids: A Strategy for Computational Screening of Electrolytes for Lithium Metal Batteries,” Computational & Theoretical Chemistry 1252 (2025): 115379, 10.1016/j.comptc.2025.115379. [DOI] [Google Scholar]
  • 37. Jain A., Ong S. P., Hautier G., et al., “Commentary: The Materials Project: A Materials Genome Approach to Accelerating Materials Innovation,” APL Materials 1 (2013): 011002, 10.1063/1.4812323. [DOI] [Google Scholar]
  • 38. Wong F., Zheng E. J., Valeri J. A., et al., “Discovery of a Structural Class of Antibiotics With Explainable Deep Learning,” Nature 626 (2024): 177–185, 10.1038/s41586-023-06887-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39. Karlov D. S., Sosnin S., Tetko I. V., and Fedorov M. V., “Chemical Space Exploration Guided by Deep Neural Networks,” RSC Advances 9 (2019): 5151–5157, 10.1039/C8RA10182E. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Li H. B., Hao J. N., and Qiao S. Z., “AI‐Driven Electrolyte Additive Selection to Boost Aqueous Zn‐Ion Batteries Stability,” Advanced Materials 36 (2024): 2411991, 10.1002/adma.202411991. [DOI] [PubMed] [Google Scholar]
  • 41. Zhou G., Gao Z., Ding Q., et al., “Uni‐Mol: A Universal 3D Molecular Representation Learning Framework,” in The Eleventh International Conference on Learning Representations (ICLR) (2023), https://openreview.net/forum?id=6K2RM6wVqKu.
  • 42. Lu S. Q., Gao Z. F., He D., Zhang L. F., and Ke G. L., “Data‐driven Quantum Chemical Property Prediction Leveraging 3D Conformations With Uni‐Mol+,” Nature Communications 15 (2024): 7104, 10.1038/s41467-024-51321-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43. Xu J. J., Zhang J. X., Pollard T. P., et al., “Electrolyte Design for Li‐ion Batteries under Extreme Operating Conditions,” Nature 614 (2023): 694–700, 10.1038/s41586-022-05627-8. [DOI] [PubMed] [Google Scholar]
  • 44. Bender A. and Glen R. C., “Molecular Similarity: A Key Technique in Molecular Informatics,” Organic & Biomolecular Chemistry 2 (2004): 3204, 10.1039/b409813g. [DOI] [PubMed] [Google Scholar]
  • 45. Maggiora G., Vogt M., Stumpfe D., and Bajorath J., “Molecular Similarity in Medicinal Chemistry,” Journal of Medicinal Chemistry 57 (2014): 3186–3204, 10.1021/jm401411z. [DOI] [PubMed] [Google Scholar]
  • 46. Niu M., Dong L. W., Chen X. Y., et al., “Approaching Fast Ion Transport via Anion–dipole Interaction in Weakly Solvated Electrolytes Enables Stable Li‐plating Chemistry,” National Science Review 12 (2025): nwaf065. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47. Seo D. M., Boyle P. D., Borodin O., and Henderson W. A., “Li+ Cation Coordination by Acetonitrile—Insights from Crystallography,” RSC Advances 2 (2012): 8014–8019, 10.1039/c2ra21290k. [DOI] [Google Scholar]
  • 48. Ugata Y., Shigenobu K., Tatara R., Ueno K., Watanabe M., and Dokko K., “Solvate Electrolytes for Li and Na Batteries: Structures, Transport Properties, and Electrochemistry,” Phys Chem ChemPhys 23 (2021): 21419–21436, 10.1039/D1CP02946K. [DOI] [PubMed] [Google Scholar]
  • 49. Xue W. J., Gao R., Shi Z., et al., “Stabilizing Electrode–electrolyte Interfaces to Realize High‐voltage Li||LiCoO2 Batteries by a Sulfonamide‐based Electrolyte,” Energy & Environmental Science 14 (2021): 6030–6040, 10.1039/D1EE01265G. [DOI] [Google Scholar]
  • 50. Bai P. X., Ji X., Zhang J. X., et al., “Formation of LiF‐Rich Cathode‐Electrolyte Interphase by Electrolyte Reduction,” Angewandte Chemie International Edition 61 (2022): e202202731, 10.1002/anie.202202731. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51. Lu T. and Chen F. W., “Multiwfn: A Multifunctional Wavefunction Analyzer,” Journal of Computational Chemistry 33 (2012): 580–592, 10.1002/jcc.22885. [DOI] [PubMed] [Google Scholar]
  • 52. Lu T., “A Comprehensive Electron Wavefunction Analysis Toolbox for Chemists, Multiwfn,” Journal of Chemical Physics 161 (2024): 082503, 10.1063/5.0216272. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Detailed information on dataset construction, DFT and molecular dynamics calculations, molecular representation, t‐SNE and clustering analysis, ESPratio validation, Uni‐Mol fine‐tuning, model evaluation, baseline comparisons, error‐propagation analysis, candidate annotation, experimental follow‐up status, and supplementary electrochemical and materials characterization data is provided in the Supporting Information. The authors have cited additional references within the Supporting Information [20, 32, 51, 52].

Supporting File: anie73524‐sup‐0001‐SuppMat.docx.

ANIE-65-e6825619-s001.docx (5.3MB, docx)

Data Availability Statement

The data and code that support the findings of this study are available at https://github.com/Hank3824/li_electrolyte_molecule_screen and https://github.com/Hank3824/Reading-notes-code/tree/main/paper%202. The first repository contains the cleaned code, curated datasets, screening outputs, candidate annotations, commercially available candidate list, the 36 previously unexplored candidates, and experimental‐status tables. The second repository contains the Materials Project and PubChem API data‐retrieval scripts used to construct the upstream training and screening datasets. Additional information is available from the corresponding author upon reasonable request.

Articles from Angewandte Chemie (International Ed. in English) are provided here courtesy of Wiley

RESOURCES