Accelerating amine-based CO₂ capture with machine learning: From molecular screening to process optimization

Abstract

Amine-based CO₂ capture represents the most mature approach for large-scale carbon reduction, with systems implemented across multiple industrial demonstration projects globally. However, vast chemical spaces encompassing millions of potential formulations and complex multiscale coupling effects pose unprecedented challenges for traditional experimental methods. Machine learning applications have achieved revolutionary advances through differentiated strategies. In liquid amine systems, ensemble learning algorithms delivered breakthrough precision improvements from traditional 4–5% to below 0.93%, while interpretable models revealed that nitrogen atom charge distribution contributes 56% to reaction barriers, enabling rational biphasic solvent design (DETA/DEEA system) that achieved 34% regeneration energy reduction compared to benchmark MEA. For solid amine systems, differential descriptor methods overcame severe overfitting challenges, improving test set performance from R² = 0.5102 to 0.79. Virtual screening of 1.6 million binding sites from the GDB-17 database identified 11% of candidates with stronger CO₂ binding than the industrial benchmark BPEI (−0.04 eV). Among these high-performance candidates, 2642 molecules simultaneously satisfied synthesizability criteria (SAscore < 3.4, GDBscore > 0.64), demonstrating both favorable binding energetics and high experimental feasibility. Critically, mechanistic analysis revealed that support physical properties dominate adsorption performance over amine chemical characteristics, fundamentally transforming material design concepts. Industrial applications demonstrated 35.76% cost reductions through intelligent solvent selection and 15–25% profit improvements through dynamic capture-level optimization combined with market-responsive bidding strategies. Despite these breakthroughs, systematic limitations, including model generalization difficulties, cross-scale integration challenges, and data standardization, persist, requiring physics-constrained algorithms and unified modeling frameworks for laboratory-to-industrial translation. These developments establish machine learning as the core driving force transitioning amine-based CO₂ capture from empirical development toward intelligent design paradigms.

1. Introduction

Amine-based technologies have emerged as the most mature approach for CO₂ capture, developing along two complementary pathways-liquid absorption and solid adsorption [[1], [2], [3], [4], [5]]. Liquid amine absorption, currently the most established technology, has been implemented in multiple million-ton-scale demonstration projects worldwide [6,7]. However, this approach still faces challenges, including high regeneration energy consumption, amine emissions, and oxidative degradation [[8], [9], [10], [11]]. Solid amine adsorbents, created by immobilizing amine molecules onto porous supports, effectively enhance amine stability while reducing regeneration energy requirements due to the absence of water [5,12,13]. Despite the distinct advantages of each pathway, both confront the same fundamental challenge: identifying optimal materials from vast chemical spaces encompassing thousands of amine molecules and countless formulation combinations [12,[14], [15], [16]].

This chemical diversity presents virtually unlimited possibilities for amine-based capture materials. In liquid amine systems, researchers must screen optimal solvents from known amine molecules, and considering mixed solvent combinations, the number of potential formulations reaches millions [17]. In solid amine systems, the complexity intensifies further: support types alone encompass hundreds of options, such as mesoporous silica, metal-organic frameworks, and porous polymers, while combinations of amine functionalization methods with supports generate thousands of possibilities [[18], [19], [20], [21], [22]]. Traditional trial-and-error approaches appear inadequate when confronting such vast parameter spaces. More critically, the performance of amine-based CO₂ capture systems exhibits nonlinear responses to operating parameters including temperature, pressure, and CO₂ concentration, with complex synergistic effects between these parameters, transforming system optimization into a high-dimensional, multi-objective problem with strong coupling effects [[23], [24], [25], [26]].

To address these challenges, the introduction of machine learning is fundamentally transforming the research paradigm in this field. In material screening, data-driven approaches have expanded evaluation scales from dozens in laboratories to millions in silico, compressing material discovery cycles from years to months [27]. For mechanistic understanding, machine learning has revealed deep patterns that traditional methods struggle to uncover. Analysis of solid amine adsorbents demonstrated that amine loading exhibits complex nonlinear relationships with CO₂ adsorption capacity: when amine loading exceeds critical thresholds relative to support pore volume, excess amine molecules obstruct pore channels and subsequently reduce mass transfer efficiency [15]. In process optimization, machine learning surrogate models have reduced process simulation time from thousands of hours to minutes while maintaining prediction accuracy, enabling near real-time process optimization. These breakthrough applications are transitioning amine-based CO₂ capture technology from an experience-driven to a data-driven new era.

This Review stands at the intersection of machine learning and amine-based CO₂ capture technology, systematically examining recent advances in this interdisciplinary field. We establish a unified analytical framework of “prediction and screening-mechanistic analysis-process optimization,” providing the first comprehensive comparison of differentiated machine learning application strategies in liquid and solid amine systems. Through analysis of successful cases, we reveal the fundamental distinction between liquid amine systems’ focus on chemical reaction mechanism modeling and solid amine systems’ emphasis on physical structure optimization. Notably, this Review not only summarizes technological achievements but also explores key challenges, including data quality, model generalization, and industrial translation, proposing solutions based on cutting-edge advances such as physics-informed neural networks and active learning. By systematically synthesizing lessons learned from machine learning-empowered amine-based CO₂ capture, this work aims to provide researchers with complete methodological guidance from molecular design to system integration, driving this critical emission reduction technology toward efficient, economical, and scalable development.

2. Machine learning-driven advancements in liquid amine-based CO₂ capture

2.1. Machine learning foundations for liquid amine systems

Liquid amine CO₂ capture systems provide an ideal technical environment for successful machine learning applications, with this compatibility stemming from the high degree of matching between system characteristics and algorithmic advantages.

Liquid amine systems possess three key characteristics that make them ideal targets for machine learning applications. (1) Data richness. This field has accumulated substantial standardized experimental data covering different amine molecules, operating conditions, and performance indicators, providing sufficient high-quality data foundations for algorithm training. (2) Behavioral complexity. System performance is influenced by nonlinear interactions among multiple factors, including temperature, pressure, amine concentration, and molecular structure, which traditional linear models struggle to accurately describe. (3) Objective diversity. Practical applications require seeking balance among multiple constraining objectives such as absorption capacity, regeneration energy consumption, reaction rates, and equipment corrosion [5,[28], [29], [30], [31], [32], [33], [34], [35]].

The core advantages of machine learning algorithms precisely match these characteristics: powerful nonlinear modeling capabilities can capture complex variable interactions, multi-output prediction frameworks can simultaneously handle multiple performance indicators, and the ability to automatically learn patterns from data eliminates the need for preset complex mathematical models.

Based on this compatibility relationship, machine learning has formed three core application directions in liquid amine CO₂ capture, each with specific technical requirements and implementation pathways. As illustrated in Fig. 1, performance prediction and material screening rely on algorithmic regression capabilities, establishing mappings between molecular structural features and macroscopic performance to achieve the leap from limited experimental data to large-scale virtual screening. Mechanism understanding and relationship analysis utilize algorithmic feature extraction and interpretability techniques to identify key influencing factors from complex data, revealing structure-performance relationships that traditional methods struggle to discover. Process optimization and intelligent control combine algorithmic multi-objective optimization capabilities to find optimal operating parameters under complex constraints, achieving global system performance optimization.

Fig. 1.

Machine learning framework for liquid amine CO₂ capture development: An integrated approach from material screening to performance prediction and system optimization.

2.2. Performance prediction and material screening

Machine learning applications in liquid amine CO₂ capture have followed a clear technological trajectory from conceptual validation to engineering implementation, witnessing not only order-of-magnitude improvements in prediction accuracy but also establishing a technological bridge from limited experimental data to large-scale material discovery.

2.2.1. Breakthrough advances in prediction precision

The pioneering breakthrough in machine learning applications emerged from directly addressing the limitations of traditional thermodynamic models in liquid amine systems. Chen et al.’s [36] groundbreaking work first demonstrated the tremendous potential of neural networks in amine-based CO₂ capture prediction. Through systematic comparison of BPNN and RBFNN across 12 single amine systems, BPNN achieved average absolute relative errors of 4.31%, 4.89%, and 3.96% for MEA, DEA, and MDEA systems, respectively, significantly outperforming RBFNN’s 4.93%, 5.17%, and 5.49%. This superior performance can be attributed to BPNN’s error back-propagation mechanism, which enables iterative weight optimization through gradient descent, effectively capturing the complex nonlinear relationships in amine-CO₂ absorption equilibria. Moreover, BPNN’s multilayer feedforward architecture achieved this with fewer hidden neurons than RBFNN’s radial basis structure, demonstrating both computational efficiency and prediction accuracy. This work established neural network superiority over eight traditional thermodynamic models, marking a paradigm shift by reducing prediction errors from the conventional 10–20% range to below 5%.

Early applications of machine learning techniques have evolved from simple models to sophisticated algorithms, successfully improving prediction accuracy for key parameters such as CO₂ solubility and mass transfer coefficients. The emergence of algorithm-task matching as a key factor for performance enhancement has been systematically validated, with different property predictions showing distinct algorithmic preferences-a discovery that provided important methodological guidance for subsequent optimization efforts [37,38].

The achievement of engineering-grade precision marked a critical transition from conceptual validation to practical application. The introduction of ensemble learning methods brought about a qualitative leap in prediction accuracy. Ghiasi et al.’s [39] AdaBoost-CART model achieved a breakthrough 0.51% prediction error in MEA systems (R² = 0.9987), with 2.76% (R² = 0.9977) and 1.41% (R² = 0.9998) for DEA and TEA systems, respectively. Fig. 2 clearly demonstrates the significance of this advancement through performance comparison among four algorithms. AdaBoost-CART predictions cluster tightly along the diagonal line (AARD = 0.93%), while LSSVM, ANN, and ANFIS all exceed 15% AARD, with particularly pronounced deviations in high CO₂ loading regions. This precision revolution established the realistic possibility of computational prediction replacing partial experimental testing, thereby laying technical foundations for large-scale virtual screening.

Fig. 2.

Comparison of ML model performance for CO₂ solubility prediction in piperazine solutions: (a) AdaBoost-CART, (b) LSSVM, (c) ANN, and (d) ANFIS approaches (reproduced from ref. [39]).

Subsequent developments have systematically validated machine learning’s adaptability to engineering complexity. Mixed amine systems, with their complex synergistic effects and competitive mechanisms far beyond simple superposition, have been successfully handled through various algorithmic approaches. Research across multiple system configurations has revealed important design principles: model complexity must match data characteristics rather than simply increasing parameters, while algorithms can automatically identify optimal concentration ratios and synergistic thresholds among different amine molecules. Comprehensive validation studies of neural network models have demonstrated remarkable stability across diverse operating conditions, maintaining high prediction accuracy for CO₂ solubility in both binary MDEA/PZ and ternary MEA/MDEA/PZ systems throughout temperature ranges of 303–393 K and pressure ranges spanning 0.06–4978 kPa. This robust performance across such wide operational windows confirms the models’ reliability for industrial applications. These advances have provided quantitative guidance for rational mixed solvent design, driving the field’s transition from empirical formulation to data-driven approaches [[40], [41], [42], [43]].

The expansion of machine learning applications to online monitoring represents another significant advancement, providing technical foundations for dynamic process control. Through integration of spectroscopic techniques with intelligent algorithms, real-time high-precision prediction of key parameters has become achievable. Syauqi et al. [44] developed an FTIR-machine learning integrated framework that demonstrates this capability’s potential. Through mutual information scoring, they selected the 950–1050 cm^-1 band (C—NH₂ vibration) for MEA concentration prediction and the 1450–1650 cm^-1 band (COO^-1 stretching) for CO₂ concentration prediction. Fig. 3 shows impressive predictive calibration results, with only 24 training samples, MEA prediction achieved R² = 0.9964 (RMSE = 0.0007), and CO₂ prediction reached R² = 0.9880 (RMSE = 0.0017). This high-precision prediction with small sample sizes opens new avenues for online monitoring, paving the way for machine learning’s evolution from offline optimization to online intelligent control.

Fig. 3.

Predictive calibration for (a) MEA concentration and (b) CO₂ loading determination (reproduced from ref. [44]).

2.2.2. The virtual screening revolution

The true transformation in material discovery emerged from virtual screening technologies, representing a paradigmatic shift from limited experimentation to large-scale database exploration. Traditional amine solvent development faced severe constraints from experimental inefficiency, confronting enormous costs and low success rates when dealing with thousands of potential molecular combinations. Machine learning fundamentally changed this landscape by establishing direct mapping relationships between molecular structural features and macroscopic performance.

Orlov et al.’s [14] pioneering work exemplifies this transformation’s potential; through QSPR models, they achieved virtual performance evaluation of over 800 commercial amine molecules from the ZINC database, compressing what would traditionally require years of experimental validation into days of computational time. Fig. 4 illustrates this high-throughput virtual screening pipeline, showing the complete integration of molecular dynamics simulations with ML-based performance prediction for tertiary amines. The practical value of this efficiency improvement is evidenced by the high experimental validation success rate: among eight candidates tested, 62.5% outperformed the benchmark material, demonstrating virtual screening’s real-world utility.

Fig. 4.

High-throughput virtual screening pipeline: Integration of molecular dynamics simulations with ML-based performance prediction for tertiary amines (reproduced from ref. [14]).

Further developments in screening strategies have expanded to stability assessment and multi-stage optimization applications. Machine learning models can now predict amine oxidative stability without time-consuming degradation experiments, discovering quantitative structural stability patterns, while multi-stage screening approaches have achieved order-of-magnitude efficiency improvements, reducing detailed evaluation workloads by over 80% through intelligent pre-screening strategies. Fig. 5 shows experimental validation comparing MLR versus CatBoost models for 27 test compounds, with CatBoost achieving AAD = 0.3% prediction accuracy [11].

Fig. 5.

Oxidative stability prediction of amine solvents: Experimental validation of MLR versus CatBoost models for 27 test compounds (reproduced from ref. [11]).

These developments collectively demonstrate that machine learning has evolved from an auxiliary tool to the core driving force for innovation in liquid amine CO₂ capture technology. Beyond achieving revolutionary improvements in prediction accuracy, it has established a complete data-driven framework from molecular design to engineering applications, providing powerful technical support for the technology’s transition to large-scale industrial deployment.

2.3. Revealing molecular mechanisms through interpretable models

The progression from high-precision prediction to mechanistic understanding represents a critical transformation in liquid amine CO₂ capture research. While achieving prediction errors below 1% marks significant technical progress, the deeper challenge lies in understanding the underlying chemical principles that govern system performance. This transition from black-box prediction to interpretable insights directly impacts the feasibility of rational solvent design and process optimization.

2.3.1. Deep analysis of molecular structural features

Unlike solid amine systems that primarily focus on physical structure optimization, liquid amine systems demand comprehensive understanding of complex chemical reaction networks and molecular interaction mechanisms. Interpretable machine learning techniques-particularly feature importance analysis, SHAP value calculation, and reaction pathway visualization-have emerged as powerful tools for unraveling these chemical complexities, transforming algorithmic outputs into actionable chemical knowledge [16,45].

Shen et al. [46] demonstrated this approach through systematic evaluation of 30 activators in biphasic systems using XGBoost models combined with SHAP analysis and DFT calculations at B3LYP-D3/6-311++G(d,p) level. Their work revealed that ADCH charge on nitrogen atoms serves as the most critical factor affecting reaction barriers, contributing 56% to model predictions (Fig. 6a). The quantum chemical calculations elucidated the underlying mechanism: for primary amines with ADCH values between −0.890 and −0.865, stronger electronegativity enhances nucleophilic attack on CO₂, reducing zwitterion formation barriers from 30 to 20 kJ/mol; conversely, for secondary amines, excessive electronegativity (ADCH < −0.88) stabilizes carbamate intermediates, hindering deprotonation and increasing barriers to 30–40 kJ/mol, as confirmed by the break point observed in ADCH-Gibbs energy relationship in Fig. 6b.

Fig. 6.

Mechanistic insights from ML analysis: (a) Feature importance ranking for reaction energy barriers and (b) ADCH charge-energy barrier correlation revealing primary/secondary amine transition (reproduced from ref. [46]).

Based on these mechanistic insights, the research team distilled five core design principles: pKa control (9.5–10.5), molar refractivity thresholds (40–50), moderate hydrophobicity (LogP 0-2), optimal ADCH range (−0.88 to −0.86), and surface tension limits (< 35 dyne/cm). Experimental validation proved this rational approach-the DETA/DEEA system, selected following these principles, achieved regeneration heat of only 2.44 GJ/t CO₂, 34% lower than benchmark MEA. This success demonstrates how mechanistic understanding transforms material development from empirical trial-and-error to theory-guided design.

Quantitative understanding of molecular surface charge distribution effects on phase behavior has been further enhanced through advanced interpretable techniques. Research combining quantum chemical calculations with machine learning revealed that ion pairs formed after CO₂ absorption exhibit broader charge distributions, significantly increasing immiscibility with organic solvents. This provides clear molecular design principles for efficient biphasic absorption systems. Additionally, analysis of mixed amine systems showed negative SHAP values for temperature, revealing its adverse effect mechanisms on CO₂ solubility, while positive SHAP values for CO₂ partial pressure validated Henry’s law applicability even in complex amine systems [[47], [48], [49]].

2.3.2. Deep understanding of mass transfer-reaction coupling mechanisms

Innovative combinations of machine learning and engineering analysis have revealed deep insights into mass transfer-reaction coupling mechanisms. Zafari and Ghaemi [50] pioneered a dimensionless analysis approach using the Buckingham Pi theorem to transform complex multi-variable systems into interpretable mechanistic insights. By reducing 11 physical parameters to seven dimensionless groups, they exposed previously hidden coupling effects: the film conversion parameter M revealed critical competition between reaction kinetics and diffusion limitations, while the interaction between loading (α) and pressure ratios (P_CO2/P_t) demonstrated how chemical equilibrium governs the available driving force for mass transfer. This mechanistic framework enabled the identification of optimal operating windows where reaction enhancement and mass transfer resistance achieve a favorable balance. Through systematic exploration of these dimensionless correlations, the study achieved optimization of CO₂ absorption performance, with the mass transfer flux reaching 0.01923 mol/m²·s under conditions that balanced reaction enhancement (M = 408.593) with transport limitations (δ_G/δ_L = 108.430).

Systematic understanding of multi-field coupling of heat transfer-mass transfer-reaction in industrial processes has been achieved through neural network sensitivity analysis [51], capturing temperature bulging phenomena at absorber bottoms that directly reflect the strong exothermic nature of CO₂ absorption and its feedback effects on mass transfer and reaction processes. Ionic speciation analysis further revealed reaction pathway complexity. At low CO₂ loading, strong alkalinity makes CO₃^2- dominant, while as loading increases, CO₃^2- transforms to HCO₃^- through protonation. This ionic equilibrium transition explains the limitations of simple kinetic models. Systematic analysis of reaction pathway complexity indicates that accuracy differences between traditional base-catalyzed hydration mechanisms (AAD = 10%) and neural network models (e.g., BPNN with AAD = 1.77% and RBFNN with AAD = 2.60%) suggest the existence of additional reaction pathways [52].

2.3.3. Mechanism-driven precision control for environmental impacts

Systematic expansion of mechanistic understanding dimensions further demonstrates the value of interpretable machine learning. Jablonka et al. [53] innovatively applied machine learning to environmental impact analysis of carbon capture processes, focusing on amine emissions as a critical environmental indicator rather than conventional CO₂ capture efficiency. They constructed a causal impact analysis framework based on gradient-boosted decision trees, quantitatively assessing how operational parameter changes affect CESAR1 mixed solvent amine emissions through counterfactual baselines. Fig. 7 shows the results of this machine learning-driven environmental impact assessment, clearly displaying temperature-dependent emission patterns for AMP and piperazine in industrial carbon capture operations. The research revealed an important discovery: when lean solvent temperature increased, the two amine components exhibited opposite emission responses-AMP emissions increased with temperature, consistent with classical volatility mechanisms, while piperazine emissions significantly decreased, suggesting predominant aerosol-based emission pathways. Further sensitivity analysis uncovered complex trade-offs among operational parameters. Reducing water wash and lean solvent temperatures decreased AMP volatile emissions but increased piperazine aerosol emissions; conversely, certain temperature combinations where top bed temperature increased actually favored piperazine emission reduction. Quantitative analysis showed AMP partial pressure exceeded piperazine by approximately two orders of magnitude, explaining why traditional unified control strategies fail in mixed amine systems. This mechanistic breakthrough demonstrates that effective emission control requires differentiated strategies based on distinct physicochemical properties of different amine components, marking a transition from empirical operation to mechanism-driven precision control.

Fig. 7.

Machine learning-driven environmental impact assessment: Temperature-dependent emission patterns for AMP and piperazine in industrial carbon capture operations (reproduced from ref. [53]).

These advances collectively demonstrate that interpretable machine learning is constructing a complete mechanistic understanding system from molecular features to macroscopic performance. Through technical approaches such as SHAP analysis, feature importance assessment, and causal inference, researchers can not only predict “which materials are good” but also understand “why they are good” and “how to make them better,” truly achieving the transformation from empirical trial-and-error to rational design. This cognitive revolution provides solid scientific foundations for continued innovation and optimization of liquid amine CO₂ capture technology.

2.4. Engineering applications and process optimization

High-precision prediction capabilities and deep mechanistic understanding now enable the translation of scientific insights into practical engineering optimization and tangible economic benefits for liquid amine CO₂ capture technology. This transition from “knowing what” and “understanding why” to “how to do better” represents a critical transformation from scientific research to engineering applications.

2.4.1. Intelligent optimization of operating conditions

Liquid amine CO₂ capture processes involve numerous interconnected operating variables, with traditional optimization methods proving inadequate when facing such multi-variable, strongly coupled, nonlinear complex systems. System efficiency is synergistically influenced by multiple parameters, including temperature, pressure, flow rates, and amine concentrations, with complex trade-off relationships among these parameters-improving one performance indicator often comes at the expense of others. Adding further complexity, industrial applications must balance multiple constraints spanning economics, safety, and environmental impacts. The introduction of machine learning methods provides entirely new technical pathways for solving this optimization complexity through constructing efficient surrogate models, achieving multi-objective collaborative optimization, supporting real-time dynamic control, and driving the field’s transformation from traditional empirical operation toward intelligent precision control.

Traditional operating condition optimization relies heavily on time-consuming simulation calculations, struggling to achieve global optima with slow response times. Breakthrough surrogate model technology effectively addresses this bottleneck by constructing artificial neural network-based substitute models to replace computationally intensive mechanistic models. Shalaby et al. [54] constructed an ANN-based surrogate model to replace the computationally intensive gPROMS mechanistic model. Through systematic parameter scanning, they generated 324 datasets covering six key variables, including reboiler duty (141–170 MW), condenser duty (6.6–10.6 MW), and flue gas flow rate. The trained ANN model not only achieved exceptional prediction accuracy (R² > 0.98) but, more importantly, reduced optimization time from hours to seconds. Building on this foundation, they compared two optimization strategies, Sequential Quadratic Programming (SQP) identified the optimal balance between energy consumption and performance (10,415.93 kJ/kg CO₂, 97.073% capture rate), while Genetic Algorithm (GA) achieved a higher capture rate of 99.696% but at the cost of a 43.4% increase in energy consumption, revealing the energy efficiency penalty of pursuing excessively high capture rates.

Deep application of parameter sensitivity analysis further enhances optimization precision and reliability. Ashraf and Dua [55] developed a partial derivative-based sensitivity analysis framework, quantitatively evaluating the influence weights of operating variables, discovering that lean solvent temperature contributed 36.6% of the impact, lean solvent flow rate accounted for 19%, while flue gas temperature contributed only 4%. Fig. 8 presents convergence behavior analysis of ANN-integrated optimization under different initial guesses, with optimization algorithms converging to similar operating regions despite starting from different points, demonstrating solution robustness. More significantly, optimal values do not always occur at variable boundaries, lean solvent temperature stabilized within a narrow range of 316 ± 7.53K, MEA concentration at the middle value of 0.29 ± 0.03 within the 0.21–0.35 range, and inlet flue gas pressure at only 0.56 ± 0.3 Pa (< 1.06 Pa). This “moderation principle” avoids unnecessary extreme operations, providing important guidance for practical engineering applications.

Fig. 8.

Convergence analysis of ANN-integrated optimization: Multiple initial guess trajectories leading to optimal MEA-based capture conditions (reproduced from ref.[55]).

Applications in process intensification equipment optimization further validate the broad applicability of intelligent optimization methods. Optimization of novel equipment such as microfluidic reactors and membrane contactors has demonstrated machine learning’s capability in handling complex mass transfer enhancement processes, with systematic investigation of solvent concentration, temperature, and flow rate effects on CO₂ removal efficiency, mass transfer coefficients, and mass transfer flux through a combination of response surface methodology and artificial neural networks. The optimized MLP model achieved 92.24% removal efficiency and 155.12 kmol/m³·hr·kPa mass transfer coefficient under optimal conditions (30 wt% concentration, 40 °C, 9 mL/min), demonstrating the tremendous potential of combining process intensification with intelligent optimization [56].

2.4.2. Systematic breakthroughs in multi-objective collaborative optimization

Actual industrial processes often require seeking balance among multiple conflicting objectives, with complexity far exceeding traditional single-objective optimization method capabilities. The introduction of advanced optimization technologies, such as Non-dominated Sorting Genetic Algorithm (NSGA-II), effectively addresses this challenge. Tikadar et al. [57] systematically studied multi-objective optimization challenges in natural gas sweetening process retrofitting, employing the NSGA-II algorithm to simultaneously optimize three objectives, minimize damage index (DI), minimize payback period (PBP), and maximize CO₂ removal rate. They developed two ProMax models for the base case and lean vapor compression (LVC) retrofit case, integrating the optimization algorithm with process simulation through an Excel-VBA interface. Fig. 9 illustrates the results of ranking Pareto optimal solutions using the TOPSIS method, clearly presenting the stratified distribution of solutions at different quality levels (best 5%, 10%, 15%, 20%, and bottom 50%). This visualization not only helps decision-makers understand solution quality distribution but, more importantly, reveals trade-offs between objectives. LVC retrofitting increased CO₂ removal from 93.64% to 97.98%, but extended the payback period from 0.61 to 0.83 years while maintaining an acceptable safety level (DI = 476).

Fig. 9.

Multi-objective optimization landscape: TOPSIS-ranked pareto solutions balancing (a) Safety-economics, (b) Performance-economics, and (c) Performance-safety trade-offs in gas sweetening retrofit (reproduced from ref. [57]).

Superstructure optimization technology development further advances multi-objective optimization to higher complexity levels. Vo et al. [58] advanced multi-objective optimization to greater complexity through their superstructure optimization study. They constructed a comprehensive superstructure containing 12 process configurations, encompassing enhancement options such as pump-around absorber (PA), rich solvent recycle (RSR), split-flow rich solvent (SRS), and lean vapor compression (LVC). Facing the challenge of mixed optimization with discrete design choices and continuous operating variables, they innovatively employed a one-hot vector deep neural network (OHV-DNN), encoding design names as four-bit binary vectors. The model trained on 4220 ProMax simulation data achieved a 1200-fold computational acceleration. Under a 90% CO₂ capture rate constraint, the optimized MEA/PZ process, compared to conventional MEA design, reduced heating energy consumption by 30.82%, capture cost by 24.39%, while maintaining amine emissions below 1 ppm, achieving comprehensive improvements in economic, environmental, and technical performance.

Recent industrial applications further validate the practical value of machine learning in post-combustion capture system optimization. Hosseinifard et al. [59] conducted a comprehensive evaluation of amine solvent selection for the Abadan power plant using random forest models, comparing the economic performance of MEA and DGA solvents. Their results demonstrated that DGA reduced total utility system costs by 35.76% compared to MEA due to its lower solvent regeneration requirements. This significant cost optimization showcases the potential of intelligent solvent selection strategies in industrial-scale carbon capture applications, representing another dimension of systematic optimization that complements process configuration and operational strategy improvements.

2.4.3. Experimental optimization and dynamic control

Dynamic control and real-time optimization technology development represent an important transition from static optimization to intelligent response. Traditional fixed operating strategies cannot adapt to dynamically changing environmental and market conditions, with this limitation particularly pronounced in carbon market environments. Li et al. [60] pioneered the introduction of reinforcement learning to operational optimization under carbon market conditions, developing a dynamic decision framework based on the Sarsa algorithm. They simultaneously considered five discrete capture levels (50−90%) and 16 bidding strategies, discretizing the continuous state space into 560 aggregated states through state aggregation techniques. After 100 quarters (25 years) of simulated learning, the agent mastered flexible market response strategies, reducing capture rates to 50–60% during low carbon price periods to minimize costs, and increasing to 80–90% during price upswings to maximize carbon credit revenue. This dynamic strategy improved cumulative discounted profit by 15–25% compared to fixed capture rate strategies, opening new directions for intelligent operation of carbon capture systems.

Intelligent development of experimental optimization technologies further demonstrates the transformational value of machine learning in engineering practice. Traditional experimental optimization methods are inefficient and costly, particularly pronounced when facing complex problems like novel amine solvent development. Hwang et al. [61] introduced Gaussian Process Bayesian Optimization (GPBO) to the development of the novel water-lean amine solvent K₂Sol, demonstrating the revolutionary role of machine learning in experimental design. The core advantage of GPBO lies in active learning, updating the probabilistic model after each experiment and achieving a balance between exploring unknown regions and exploiting known information through the Thompson sampling strategy. With fewer than 30 experiments, they determined K₂Sol’s optimal operating conditions (L/G ratio 5.67 L/Nm³, stripper pressure one bar), achieving an ultra-low regeneration energy of 2.8 GJ/t CO₂, only 65% of conventional MEA. More importantly, K₂Sol maintained excellent performance after five absorption-desorption cycles, while MEA performance degraded by over 30%, proving the value of intelligent optimization methods in new material development.

These advances collectively demonstrate that machine learning has become the core driving force for liquid amine CO₂ capture engineering optimization. From intelligent optimization of operating conditions to multi-objective system design, from static parameter adjustment to dynamic market response, machine learning technologies have constructed complete intelligent engineering systems. This technological revolution not only significantly improves system economics and reliability but, more importantly, provides powerful methodological support for large-scale industrial deployment of this technology, driving liquid amine CO₂ capture’s fundamental transformation from traditional experience-driven to intelligent data-driven approaches.

3. Machine learning-enabled innovations in solid amine adsorbent development

3.1. Machine learning computational foundations for solid amine systems

Solid amine CO₂ capture systems exhibit markedly different technical characteristics from liquid amine systems, with these differences creating unique challenges and opportunities for machine learning applications. Solid amine technology immobilizes organic amine molecules onto porous solid supports, enabling direct CO₂ adsorption at the solid-gas interface and fundamentally resolving key issues that have long plagued liquid amine systems, including solvent handling, equipment corrosion, and high-energy regeneration processes [5].

The complexity of solid amine systems stems from their multiscale coupling characteristics. At the molecular level, different amine functional groups (primary, secondary, and tertiary amines) exhibit distinct interaction mechanisms with CO₂. At the support level, structural features including pore size distribution, specific surface area, and pore volume significantly influence overall performance. At the preparation level, different methods such as physical impregnation [62], chemical grafting [62], in-situ polymerization [63], and dry synthesis [64] exert important impacts on final performance. This multilevel complexity requires machine learning methods to simultaneously handle molecular-scale chemical interactions, material-scale physical structural features, and process-scale preparation parameters, forming comprehensive modeling capabilities across multiple scales. Unlike liquid amine systems that primarily focus on chemical reaction mechanisms, solid amine systems emphasize physical structure optimization and understanding of multiscale coupling effects.

Based on these system characteristics, as shown in Fig. 10, machine learning has formed three interconnected core application directions in solid amine CO₂ capture, material screening and design establishes structure-performance mapping relationships to progressively narrow candidate selection from large databases, achieving efficient material discovery; performance prediction and mechanism analysis provides accurate CO₂ adsorption capacity forecasting and quantified feature importance analysis, deepening understanding of influential factors; process optimization and system integration employs parameter sensitivity analysis and multi-objective optimization to seek optimal balance between energy consumption and performance. This multilevel application framework has transformed solid amine adsorbent development from an experiment-driven process to a data-driven approach, substantially enhancing research efficiency and expanding performance boundaries. The unique advantages demonstrated by machine learning methods in handling the heterogeneity, multiscale nature, and nonlinear coupling characteristics of solid amine systems provide powerful methodological support for this technology’s transition from laboratory research to industrial applications.

Fig. 10.

Comprehensive ML application strategy for solid amine adsorbents: Three-pillar framework spanning material discovery, property prediction, and process integration.

However, current machine learning research on solid amine adsorbents exhibits limitations in the systematic characterization of multi-scale parameters. Existing studies primarily represent differences among carrier types through unified physical descriptors such as pore volume, pore size, and BET surface area, while the intrinsic chemical nature of carriers-including surface hydroxyl groups on oxides, hydrophobicity of carbon materials, and coordination chemistry of MOFs-is typically reflected only indirectly through elemental composition. Similarly, preparation methods (e.g., physical impregnation versus chemical grafting) are rarely incorporated as explicit features in models, largely due to low process standardization and the dominance of physical impregnation in publicly available datasets, resulting in preparation effects being captured indirectly through final pore structure parameters. Consequently, the systematic impacts of carrier types and preparation methods on adsorption performance await elucidation through standardized data collection, feature engineering innovations, and targeted correlation analysis. The following sections discuss recent advances in machine learning for predictive modeling and mechanistic interpretation of solid amine adsorbents, with explicit recognition of these research boundaries.

3.2. Predictive modeling and high-throughput screening breakthroughs

Solid amine material discovery is undergoing a profound paradigmatic revolution, from empirical exploration based on experimental intuition toward rational design driven by data insights. The core of this transformation lies in addressing two major technical bottlenecks that have long plagued the field: the reliability crisis in predictive models and the scalability challenges in material screening.

3.2.1. Evolution of performance prediction model

Machine learning applications in solid amine CO₂ adsorption performance prediction face distinctly different challenges compared to liquid amine systems. The heterogeneity and complexity of solid amine materials impose higher demands on predictive models, driving the evolution of prediction methods from simple parameter mapping toward deep mechanistic understanding.

The research by Adewunmi et al. [65] vividly demonstrates the severity of this challenge. They developed a support vector regression model integrating six key parameters-support pore size, surface area, polymer molecular weight, polymer concentration, temperature, and pressure-to predict CO₂ adsorption capacity. On the training set of 99 data points, the model achieved near-perfect fitting with R² = 0.9999. However, as the stark contrast between Fig. 11a and b reveals, the model’s performance on the test set plummeted to R² = 0.5102, with RMSE soaring from 0.068 to 0.9291. This performance collapse stems from three fundamental factors: First, the six macroscopic descriptors fail to capture the microscopic heterogeneity of solid amine materials-local amine clustering within pores, pore blocking effects, and interfacial hydrogen bonding networks. Second, training data predominantly comprised PEI-functionalized SBA-15 and MCM-41 systems, while test sets included structurally distinct materials, causing model failure when extrapolating beyond its learned chemical space. Third, the purely data-driven SVR lacks adsorption thermodynamic and kinetic constraints, potentially learning spurious correlations rather than genuine physical laws. This case transcends conventional overfitting problems, revealing the fundamental challenge of effectively modeling highly heterogeneous materials.

Fig. 11.

SVR model performance contrast: (a) Near-perfect training set fitting versus (b) Significant test set deterioration in polymer-amine CO₂ adsorption prediction (reproduced from ref. [65]).

The genuine breakthrough emerged from a fundamental shift in modeling philosophy-from describing material states to capturing functionalization-induced changes. The differential descriptor approach, mathematically defined as Δx = x_child - x_parent, where x represents molecular descriptors before and after CO₂ binding, achieved stable test set performance (R² = 0.79) by focusing on transformation rather than absolute properties. This method captures how molecular features change during amine functionalization: for instance, the change in hydrogen bond acceptors (ΔnHBAcc) directly reflects carbamate formation, while changes in topological descriptors (ΔABCGG) reveal structural reorganization upon CO₂ binding. By encoding the functionalization process as descriptor differences rather than absolute values, the model inherently accounts for the heterogeneous baseline properties of diverse amine materials, enabling robust predictions across structurally varied systems. This transformation-focused approach represents a cognitive leap in solid amine research, moving from phenomenological description toward mechanistic understanding of the CO₂ capture process [66].

The prediction and identification of complex reaction products represent another important breakthrough in the solid amine field. The CSA (Chemical Shift Anisotropy)-machine learning coupled technique developed by Marple et al. [67] cleverly overcomes the fundamental limitations of traditional ¹³C NMR in identifying carbonyl compounds. Traditional methods are constrained by the narrow chemical shift range of carbonyl compounds (< 10 ppm), making it difficult to distinguish structurally similar products. The CSA method, by introducing multidimensional information from chemical shift tensors, expands one-dimensional chemical shift data into three-dimensional tensor space, greatly enhancing structural recognition capability. Based on a DFT calculation database of 142 carbon atoms, their trained k-nearest neighbor classification model achieved breakthrough recognition accuracy. As clearly shown in the classification results in Fig. 12, in the BPEI/γ-Al₂O₃ system, component A in the 164–166 ppm region was predicted with high probability as urea (85–92%), while component B in the 161–162 ppm region was accurately predicted as carbamate (65–75%). The predictions highly matched the experimental conditions (dry CO₂ exposure and partial regeneration at 100 °C). This method, achieving 85–95% recognition accuracy, not only provides molecular-level insights into solid amine CO₂ adsorption mechanisms but also points the direction for subsequent material optimization design.

Fig. 12.

CSA-ML product identification in BPEI/γ-Al₂O₃: Probability distributions for urea versus carbamate formation based on chemical shift analysis (reproduced from ref.[67]).

3.2.2. Computational screening frameworks for adsorbent design

The establishment of reliable predictive models has made systematic virtual screening approaches possible, fundamentally transforming material discovery paradigms. Davis et al. [66] constructed a comprehensive three-stage screening framework to evaluate over 1.6 million nitrogen-containing heterocyclic compounds from the GDB-17 database. As shown in Fig. 13, the screening results revealed that 11% of molecules exhibited stronger CO₂ binding capabilities than the industrial benchmark BPEI, indicating enormous untapped potential for high-performance materials in chemical space.

Fig. 13.

Large-scale virtual screening outcomes: (a) Binding energy distribution across 1.6M GDB-17 molecules, and (b-d) Representative structures from high, low, and neutral binding regions post-synthesizability filtering (reproduced from ref. [66]).

The screening strategy incorporated multidimensional evaluation criteria beyond CO₂ binding energy. The research team introduced synthetic accessibility scores (SAscore < 3.4) and drug-likeness scores (GDBscore > 0.64) as additional screening criteria to ensure practical synthesizability. This multi-objective optimization approach ultimately identified 2642 candidates from the high binding energy region that simultaneously met performance and synthesizability requirements. These materials exhibited remarkable structural diversity, with an average Tanimoto similarity of only 0.15, including many fused aromatic ring systems and complex heterocyclic structures that differ significantly from traditional design approaches.

Computational efficiency advantages are equally important for practical deployment. Owens et al. [68] developed a two-tier computational workflow combining fast pre-relaxation using m3GNet neural network potentials with high-precision DFT validation. This approach provides a 3–4 order-of-magnitude speed-up over conventional DFT while maintaining quantitative accuracy, enabling assessment of tens of thousands of configurations within hours rather than months. The integration of predictive modeling with experimental validation has established complete material discovery pipelines. Virtual screening approaches have demonstrated practical value through high experimental validation success rates-computational predictions successfully identified multiple candidates that outperformed benchmark materials in subsequent experimental testing. This integration of computational prediction with targeted experimental validation represents a significant efficiency improvement over traditional trial-and-error approaches, enabling researchers to focus experimental resources on the most promising candidates identified through intelligent screening algorithms.

Machine learning-accelerated material discovery methods have successfully established a complete technical chain from prediction to screening in the solid amine field. Various advanced algorithms, including deep neural networks, random forests, gradient boosting, and Bayesian optimization, have all demonstrated significant advantages in performance prediction and material screening, achieving speed improvements of three or more orders of magnitude compared to traditional molecular simulations. The successful applications of these methods mark the transition of solid amine material discovery from experiment-driven serendipitous discovery to data-driven rational design, providing powerful technical support for the industrial applications of this technology (multiple researchers, including Davis, Owens, Marple, and Adewunmi).

3.3. Mechanistic understanding through interpretable machine learning

The achievement of high-precision prediction has presented a deeper challenge for the solid amine CO₂ capture field, progressing from “knowing results” to “understanding causes.” This cognitive leap from black-box prediction to white-box insight not only concerns the completeness of scientific understanding but also directly impacts the feasibility of rational design and engineering optimization.

3.3.1. Quantitative analysis of multiscale feature influence mechanisms

The complexity of solid amine systems stems from their multiscale coupling characteristics, molecular-scale amine-CO₂ interactions, material-scale pore structure effects, and preparation-scale process parameter influences. This multilevel complexity requires researchers not only to predict “which materials perform well” but to understand “why they excel” and “how to optimize them systematically.” Unlike liquid amine systems that focus on chemical reaction mechanism modeling, solid amine systems demand quantitative understanding of multiscale feature influence mechanisms. Interpretable machine learning provides essential tools for addressing this challenge through SHAP analysis, feature importance assessment, and multiscale coupling visualization, enabling quantitative separation of different-scale features’ contributions to overall performance. This multiscale mechanistic understanding guides materials scientists toward collaborative optimization strategies that simultaneously consider support structure, amine loading, and interface interactions, thereby establishing technical foundations for rational multiscale design.

Li et al. [69] pioneered the integration of ChatGPT 4o with SHAP analysis to systematically elucidate the multi-level influence mechanisms governing solid amine adsorption performance. As illustrated in Fig. 14a, SHAP values clearly demonstrate the contribution hierarchy of various features, pore volume (PVBD) ranks first, followed by specific surface area (SSAAD) and amine loading (LA). This ranking reveals a crucial mechanism: pore volume emerges as the primary factor because it directly controls the contact efficiency between CO₂ molecules and amine groups. Larger pore volumes enable more amine groups to remain exposed on accessible surfaces. The mean SHAP value analysis in Fig. 14b further categorizes the influencing factors; physical properties of solid supports contribute over half of the total impact, followed by chemical properties of loaded amines, while testing conditions exert relatively minor influence. Particularly noteworthy is that the SHAP waterfall plots unveil the chemical reaction mechanism: when nBase = −1, it indicates consumption of one basic amine group during carbamate formation with CO₂, and this depletion of basic sites directly affects the binding enthalpy (−0.167 eV). This discovery explains why certain materials with theoretically high amine content underperform in practice-the key lies not in total amine quantity but in the number of effective amine groups available for reaction.

Fig. 14.

(a) Individual feature contributions via SHAP values and (b) Aggregated impact analysis revealing physical property dominance (reproduced from ref. [69]).

Zhang et al.’s [15] work further demonstrates how machine learning transforms mechanistic insights into practical material design principles. Through random forest models combined with SHAP analysis, this study systematically dissected the influence of five key factors on solid amine adsorption performance. The impact of amine loading on CO₂ adsorption capacity exhibits a distinct duality. SHAP value analysis reveals that smaller amine loading feature values correspond to positive SHAP values, indicating that increasing loading enhances adsorption capacity; however, when loading feature values are large, some data points’ SHAP values fall below zero, directly confirming the negative effects of excessive loading. This nonlinear mechanism stems from the delicate balance between amine loading and pore volume; moderate amounts of amine provide adsorption sites, while excess amine blocks the pore and hinders mass transfer. Based on this mechanistic understanding, the study determined optimal loading strategies for different amine types. Linear TEPA should occupy 40–50% of the support’s pore volume, while polymeric PEI requires 60–75% due to steric hindrance. More importantly, applying these quantitative structure-property relationships, the team successfully predicted the optimal formulation for the new support KYX, with 45TEPA/KYX’s predicted adsorption capacity (3.82 mmol/g) closely matching the experimental value (3.78 mmol/g), powerfully demonstrating the feasibility of a complete research paradigm from mechanistic analysis to rational design to precise prediction.

The discovery of pore confinement effects provides a theoretical foundation for understanding “unexpected winners” in screening results. Zhang et al. [15] noticed that certain materials marginalized in initial screening due to smaller pore sizes showed surprisingly high amine efficiency in actual tests. In-depth research found that the 5–10 nm pore size range forms an ideal “confined space” that can accommodate molecular motion while enhancing CO₂-amine interactions through van der Waals forces. This mechanistic understanding not only explains the “dark horse” phenomenon in screening but also provides clear guidance for subsequent targeted design. Rather than pursuing extremely large pore sizes, pore structures should be precisely matched to amine molecular dimensions.

The above SHAP-based interpretability analyses were primarily conducted on specific types of carrier materials. Li et al.’s study covered porous silica and alumina carriers, while Zhang et al. focused on commercial mesoporous silica gels and activated carbon materials. Although these studies successfully revealed the importance rankings and influence mechanisms of parameters such as pore volume, BET surface area, and amine loading in the investigated systems, the applicability of these conclusions across broader material systems requires further verification. Additionally, in current research, the influence of preparation methods is mainly reflected indirectly through final pore structure parameters, and how different preparation routes (such as physical impregnation versus chemical grafting) directly affect feature importance rankings has not been systematically quantified. Despite these research boundaries, the successful applications of interpretable machine learning techniques in the solid amine field have established a complete methodology from data mining to mechanistic understanding. Various advanced interpretation techniques, including SHAP analysis, feature importance assessment, and decision tree visualization, have all demonstrated powerful capabilities in revealing complex structure-performance relationships. These methods can not only predict “which materials are good” but also understand “why they are good” and “how to make them better,” truly achieving the transformation from empirical trial-and-error to rational design.

3.4. Multi-Objective process design and system optimization

The transition of solid amine CO₂ capture technology to practical engineering applications requires seeking an optimal balance among multiple constraining performance indicators and translating this balance into actionable engineering design guidance. Breakthroughs in machine learning methods for multi-objective optimization and system integration provide powerful technical support for this transition.

3.4.1. Multi-Objective collaborative optimization in material design

Zhang et al. [15] achieved multi-objective optimization of solid amine adsorbents through random forest models, transforming the traditional pursuit of single performance metrics into a comprehensive balance of five key indicators. The research team simultaneously considered CO₂ adsorption capacity, amine efficiency, adsorption rate, τ90, and cycling stability, using machine learning to reveal complex trade-off relationships among these indicators-for instance, while amine loading increases adsorption capacity, it reduces amine efficiency, and while larger pore sizes facilitate mass transfer, they sacrifice specific surface area. Particularly noteworthy is their discovery of parameter combination “sweet spots”, when pore volume is 0.8–1.2 cm³/g, TEPA and PEI exhibit optimal loading windows of 40–50% and 60–75% respectively, within which multiple performance indicators can simultaneously achieve favorable levels. The five-dimensional radar chart vividly illustrates these multi-objective optimization results, with 45TEPA/KYX and 60PEI/KYX displaying distinctly different performance profiles-the former excelling in adsorption performance while the latter demonstrating superior stability, proving that no absolute “optimal” material exists, only best compromise solutions for specific application scenarios.

3.4.2. Breakthrough advances in system-level engineering optimization

To design a DAC system capable of simultaneously achieving low energy consumption, high productivity, and high purity, Young et al. [26] conducted an unprecedented multi-dimensional parameter optimization study. The research team systematically evaluated the combined effects of 15–16 key parameters on five performance indicators (specific thermal energy, specific electrical energy, productivity, purity, and working capacity), encompassing material properties (CO₂ mass transfer coefficient, thermal conductivity, adsorbent density, adsorption isotherm parameters) and operating conditions (vacuum pressure, heating temperature, air flow rate). Facing the 42-year computational time required for 2.5 million simulations, they innovatively adopted a machine learning acceleration strategy, first generating 30,000 high-fidelity training samples using detailed TVSA and S-TVSA process models, then training multi-output artificial neural networks (14 layers deep, 1000 nodes per layer) to learn the complex parameter-performance mapping relationships, and finally utilizing this surrogate model to complete all Sobol sensitivity analyses within seconds. The results are striking-the total-order Sobol index for CO₂ mass transfer coefficient’s effect on specific thermal energy is three times its first-order index, this vast difference directly proving the criticality of parameter interactions and completely negating the effectiveness of traditional local optimization methods.

Through systematic analysis of 2.5 million data points, Young et al. [26] revealed the “dual-threshold” constraints on DAC system performance. Fig. 15a and 15b together outline a clear optimization roadmap, to enter the ideal operating window of low energy consumption (< 1.5 MJ/kg) and high productivity (> 200 kg/m³/day), the system must simultaneously cross two thresholds-heat transfer coefficient must exceed 20 W/m²/K (deep red region in Fig. 15a) and CO₂ mass transfer coefficient must exceed 0.003 s^-1 (orange-red region in Fig. 15b), with any single parameter deficiency locking the system in suboptimal regions. More importantly, they discovered that the maximum CO₂ adsorption capacity, which traditional research overly focuses on, has almost no impact on system performance in TVSA processes, while the overlooked material density and heat capacity have decisive effects on productivity-combinations of high density (> 1000 kg/m³) and low heat capacity (< 1000 J/kg/K) can increase productivity by over 50%. Based on these quantitative findings, the research clarified DAC material design priorities, first improving mass transfer kinetics (target kCO₂ > 0.01 s^-1) and then enhancing thermal conductivity (> 10 W/m/K), while optimizing support density and heat capacity, rather than blindly pursuing ultimate adsorption capacity. This data-driven multi-dimensional optimization framework not only provides scientific basis for current material screening but also points the direction for next-generation DAC technology development.

Fig. 15.

Process performance trade-off analysis: (a) Heat transfer coefficient impact and (b) Mass transfer coefficient influence on energy-productivity relationships in DAC systems (reproduced from ref. [26]).

The successful applications of multi-objective optimization and system integration technologies in the solid amine field mark that this technology has acquired the capability for transitioning from laboratory to industrial applications. Various advanced optimization algorithms, including multi-objective genetic algorithms, Pareto frontier analysis, and sensitivity analysis, have all demonstrated powerful capabilities in balancing complex engineering constraints. The successful applications of these methods enable solid amine CO₂ capture technology to simultaneously meet stringent industrial requirements across multiple dimensions, including performance, cost, and stability, laying solid foundations for large-scale deployment of this technology.

4. Perspectives and future horizons

4.1. Can machine learning really revolutionize amine-based CO₂ capture?

Based on systematic analysis in this review, the answer is affirmative, but this revolutionary change has clear boundaries and conditions. Machine learning has achieved qualitative breakthroughs in core aspects of amine-based CO₂ capture, from efficiency revolution in material discovery to depth advancement in mechanistic understanding, and precision enhancement in engineering optimization. Each dimension demonstrates capabilities that surpass traditional methods. Most importantly, these breakthroughs are not isolated technical advances but mutually reinforcing systematic transformations that are redefining the research paradigm of the entire field.

The differentiated development of liquid and solid amine systems demonstrates machine learning’s targeted effectiveness when methods align with system-specific challenges. Liquid amine systems achieved quantifiable success in reaction mechanism modeling and process optimization, reducing prediction errors from 10–20% to below 5% for established solvents, enabling practical industrial applications. Solid amine systems advanced through innovative approaches like differential descriptors that specifically address material heterogeneity, improving model stability despite data limitations. This complementary pattern—chemical insights for liquids, structural optimization for solids—reveals that machine learning excels not as a universal solution but as an adaptable methodology that succeeds when properly matched to domain-specific physics and chemistry.

The deeper significance lies in machine learning technology’s evolution from auxiliary tools to core driving forces of innovation. The traditional “hypothesis-verification” research mode is being supplemented or even replaced by “data-discovery” approaches, enabling researchers to automatically discover hidden patterns and mechanisms from massive datasets. This transformation in cognitive approaches not only improves research efficiency but, more importantly, expands the boundaries of human understanding of complex chemical systems, making previously intractable high-dimensional, nonlinear problems solvable.

4.2. Critical limitations and unresolved challenges

The most fundamental challenge in current applications is the reliability gap, where high-precision models under laboratory conditions often experience significant performance degradation when confronting industrial environment complexity. This degradation reflects not merely an accuracy decline but fundamental limitations of current modeling approaches, over-reliance on idealized training data and insufficient consideration of system robustness. Industrial environment factors such as temporal variability in feedstock composition, non-ideal equipment behavior, and long-term operational stability issues all exceed current model processing capabilities. Addressing this challenge requires full consideration of uncertainty and variability at the model design stage.

Data foundation fragility constitutes another systematic constraint. Unlike fields such as computer vision or natural language processing that benefit from massive, standardized datasets, amine-based CO₂ capture research relies on relatively small, heterogeneous collections of experimental measurements. Inconsistent experimental conditions, measurement method variations, and lack of systematic long-term stability studies all introduce substantial noise and bias into training data. More seriously, current lack of effective data quality control mechanisms means algorithms may learn experimental artifacts rather than true physicochemical laws, resulting in seemingly high-precision models that actually lack reliable predictive capability.

The most challenging aspect is the fundamental difficulty in feature representation, involving how to transform complex physicochemical phenomena into mathematical forms processable by algorithms. Molecular synergistic effects in liquid amines and multiscale coupling in solid amines both exhibit strong nonlinear, non-additive characteristics that existing molecular descriptors and material features struggle to capture completely. More importantly, current interpretability methods primarily reveal statistical correlations rather than causal relationships, limiting model extrapolation capability in new systems beyond training distributions. Addressing this challenge requires developing new feature engineering methods and causal inference frameworks.

An equally important limitation concerns the systematic modeling of multi-scale parameters in solid amine systems. As discussed in earlier sections, current machine learning research often simplifies carrier types and preparation methods into unified numerical descriptors during feature engineering. While this approach reduces modeling complexity, it constrains the ability to capture system-specific patterns across different material classes. The intrinsic chemical differences among carriers, such as surface hydroxyl groups on oxides versus hydrophobicity of carbon materials, and the effects of preparation routes, such as variations in amine dispersion states between physical impregnation and chemical grafting, are typically reflected only indirectly, limiting model accuracy when predicting performance of novel carrier-preparation method combinations. Overcoming this bottleneck requires establishing standardized encoding systems that preserve material chemistry information while enabling computational tractability. Simultaneously, quantitative revelation of synergistic patterns across different material systems can be achieved through correlation analysis and interaction effect elucidation, such as SHAP interaction value calculations. Furthermore, constructing hierarchical modeling frameworks-where materials are first classified by system type before establishing specialized sub-models-can enhance prediction accuracy while maintaining interpretability. Advancing these efforts necessitates not only diverse and standardized data accumulation but also sustained interdisciplinary collaboration between materials science and machine learning communities to bridge the gap between chemical intuition and algorithmic representation.

4.3. Transfer learning and cross-domain innovation pathways

Methodological breakthroughs achieved in amine-based CO₂ capture are transcending their original application boundaries, becoming significant drivers of broader technological innovation. This cross-domain impact is realized through three transfer learning modes: direct transplantation of algorithmic frameworks, adaptive improvement of feature engineering methods, and fundamental transformation of cognitive paradigms. Most representative is the differential descriptor method, which successfully addressed overfitting challenges in solid amine systems, with its core principle focusing on relative changes rather than absolute properties, providing a universal framework for any materials engineering problem involving chemical functionalization. The transferability of this methodology has been validated across multiple fields, including catalyst design and pharmaceutical materials engineering, demonstrating that intellectual tools refined from specific applications often possess unexpectedly universal value. Active learning strategies through intelligent sampling method innovations reduced required experimental data by three orders of magnitude, enabling training of high-precision predictive models with determination coefficients exceeding 0.99 using merely 0.124% of molecular samples [70].

Cross-system knowledge transfer has begun to demonstrate potential for establishing unified chemical engineering methodologies, with this transfer extending beyond algorithmic levels to embody deep understanding of common principles governing complex systems. Knowledge transfer between liquid and solid amine systems reveals deep connections between seemingly different chemical processes; for instance, understanding intermolecular interactions in liquid amines provides crucial insights for optimizing amine loading and distribution in solid amines. Interpretable machine learning techniques, particularly multiscale SHAP analysis methods, provide powerful tools for understanding complex engineering systems, with their capability to separate relative contributions of different physicochemical factors showing enormous application potential in membrane separations, heterogeneous catalysis, and advanced materials processing. More importantly, standardized database construction has achieved breakthrough progress, with the CURATED framework implementing principles of clean, uniform, refined, and automatically tracked data management, establishing complete traceable data chains from DFT-optimized structures to adsorption performance evaluation for 324 covalent organic frameworks [71]. This data standardization success establishes foundations for cross-domain materials databases, enabling truly collaborative accumulation of work from different research teams.

The most transformative development direction is the deep integration of physics-constrained machine learning frameworks with autonomous experimental systems, catalyzing entirely new modes of scientific discovery. Physics-constrained neural networks, by directly embedding two-phase flow equations into neural network loss functions, can accurately predict CO₂ pressure distribution and saturation evolution in geological reservoirs even under extremely limited training data conditions [72]. The more advanced CO₂PCNet framework, through encoder-decoder architecture combined with ConvLSTM modules, successfully achieves high-precision prediction of CO₂ plume spatiotemporal evolution in complex porous media [73]. This physics-constrained approach has expanded to multiple domains, including crystal plasticity and multi-fidelity modeling [74], fully demonstrating its universality. In amine systems, incorporating chemical absorption kinetics, thermodynamic equilibrium constraints, and mass transfer equations as intrinsic constraints of neural networks not only improves prediction accuracy but, more importantly, ensures physical reasonableness of model outputs, fundamentally resolving the limitation of purely data-driven models that “know what but not why” [75]. Genetically algorithm-optimized MOF structures demonstrated performance significantly exceeding commercial 13X zeolites in industrial-scale improved Skarstrom cycle simulations, proving the industrial viability of computationally designed materials [76]. Larger-scale validation comes from complete integration of atomic-level simulations with pilot-scale validated VSA simulators, where among 1632 experimentally characterized MOFs screened, 365 materials showed parasitic energy lower than traditional solvent-based processes, reaching as low as 217 kWhe/MT CO₂, demonstrating enormous energy-saving potential [77]. Deep fusion of machine learning with multiscale modeling has created new research paradigms capable of exploring vast design spaces while managing ill-posed problems [78]. Particularly noteworthy are the unique advantages demonstrated by graph neural networks in capturing molecular structural information and intermolecular interactions, providing direct mapping pathways from molecular structure to macroscopic performance for CO₂ solubility prediction in novel absorbents such as deep eutectic solvents [79]. Integrating multidimensional stability metrics, including thermodynamic, mechanical, thermal, and activation stability into high-throughput screening workflows, ensures that computationally screened materials possess practical application value [80].

CRediT authorship contribution statement

Ping Yang: Writing – original draft, Conceptualization. Xiaoman Yu: Writing – original draft. Kyriakos C. Stylianou: Writing – review & editing. Liang Huang: Writing – original draft, Conceptualization. Qiang Wang: Writing – review & editing, Supervision, Funding acquisition, Conceptualization.

Declaration of competing interest

The authors declare that they have no conflicts of interest in this work.

Biographies

Ping Yang is a PhD candidate at the School of Environmental Science and Engineering, Beijing Forestry University, China. She is expected to receive her PhD in 2027. Ms. Yang's research focuses on the intelligent design of solid amine CO₂ adsorbents by integrating machine learning with materials science. Her work combines physical constraints with data-driven approaches to accelerate the discovery of high-performance carbon capture materials. She has published seven papers, with four specifically addressing CO₂ capture technologies. Ms. Yang has participated in the National Key R&D Program of China and projects funded by the National Natural Science Foundation of China. She has been awarded university-level scholarships for her academic achievements.

Liang Huang (BRID:07768.00.19582) is an associate professor and master’s supervisor at the College of Environmental Science and Engineering, Beijing Forestry University, where he obtained his Ph.D. in 2019 and has since continued his research career. His primary research focuses on carbon capture, utilization, and storage (CCUS). Dr. Huang has authored over 40 SCI-indexed publications in these fields and has led or participated in several prestigious national research projects, including grants from the National Natural Science Foundation of China and the National Key R&D Program of China. His innovative work has been recognized with four authorized national invention patents and distinguished honors such as the Third Prize of Beijing Science and Technology Award and the Youth Award for Green Steel Technology Innovation.

Qiang Wang (BRID: 08557.00.50900) is a professor and vice president at Beijing Forestry University. He obtained his PhD from POSTECH (2009) after receiving bachelor’s and master’s degrees from Harbin Institute of Technology. Before joining BFU in 2012, he conducted postdoctoral research at University of Oxford after working at Singapore’s Institute of Chemical and Engineering Sciences. His research focuses on environmental catalysis and carbon capture technologies, including: (1) CO₂ capture and utilization, (2) flue gas denitration catalysts, and (3) VOC adsorption/catalytic materials. With over 260 SCI publications in journals like Angew. Chem. Int. Ed. Adv. Mater. and Energ. Environ. Sci., he has been consistently recognized as an Elsevier's Highly Cited Researcher. His honors include the National Science Fund for Distinguished Young Scholars and the Ministry of Education's Outstanding Research Award.