AI in single-atom catalysts: a review of design and applications

The initial stage: the DFT and ab initio molecular dynamics are calculated to screen the performance of the catalyst one by one

In the early stages of developing SACs, researchers primarily relied on DFT and ab initio molecular dynamics (AIMD). DFT is widely used to predict the stability, electronic structure, reaction pathways, and energy barriers of SACs^[51,52]. The catalytic activity of SACs is closely linked to the electronic structure of individual active atoms and their interactions with the support, making DFT calculations essential for simulating binding energies, adsorption energies, and reaction transition states. For example, the adsorption behavior of metal atoms on different supports, such as graphene, nitrogen-doped carbon, and titanium dioxide, can be systematically studied using DFT^[53,54]. Complementarily, AIMD integrates with DFT, simulating the spatiotemporal evolution of materials at the atomic scale to investigate the thermodynamic behavior of catalysts under realistic reaction conditions. AIMD not only validates the accuracy of DFT-optimized models but also simulates the kinetic stability of catalysts under varying temperatures and pressures, providing dynamic information crucial for predicting catalyst performance in complex reaction environments.

These computational methods can generate a wealth of structural and performance data. On the one hand, they serve as valuable empirical information; on the other hand, they provide a crucial foundation for new research. Based on existing physicochemical information, researchers can make informed predictions about the performance of similar materials and subsequently validate these predictions through experiments. This integration of theory and experimentation has driven deeper research and technological advancements in the field of SACs.

The necessity of DFT and AIMD in SACs research

SACs demonstrate tremendous application potential in the energy and environmental sectors due to their unique catalytic properties. However, traditional d-band theory struggles to explain the catalytic mechanisms of SACs, presenting ongoing challenges in theoretical characterization^[55]. To address this issue, researchers have made significant progress through the synergistic use of DFT and AIMD.

DFT can be employed to investigate the fundamental mechanisms of molecular adsorption in SACs, thus unveiling the reaction pathways. For example, He et al. investigated the catalytic oxidation of CO on SACs using DFT calculations, proposing a “selective orbital coupling” mechanism that reveals the selective coupling between the localized d orbitals of metal single atoms and the π^* orbitals of O₂ molecules^[56]. This coupling determines the strength of the M–O bond and explains the variations in energy barriers along the reaction pathway. By calculating the energies of different orbitals, they quantitatively predicted the adsorption strength and correlated it with the reaction barriers, enhancing the understanding of catalytic behavior of SACs. Using DFT calculations, researchers analyzed the highly localized d orbital characteristics of metal doping. The Wannier functions for the coupling between the d and π^* orbitals of O₂ molecules in various adsorption configurations [Figure 2A] illustrate that the most stable structure of O₂ molecules arises from the selective coupling of the π^* orbitals with specific d orbitals ($$ d_{x^2-y^2} $$) of metal dopants, forming the M–O bond. This coupling mechanism determines the strength of the M–O bond and explains the trends in reaction energy barriers during the CO oxidation reaction. Figure 2B quantitatively relates adsorption strength to reaction barriers by utilizing the energy arrangement of the selected orbitals, helping to understand how the chosen d states influence reaction behavior. DFT calculations revealed the electronic structure characteristics of SACs, providing a basis for understanding catalytic mechanisms. Our previous research also demonstrated that the sub-orbitals of d electrons play different roles in the adsorption of catalytic intermediates. For example, when surface transition metals (TM) bind with adsorbed molecules, the coupling of the 3σ → $$ d_{z^2} $$ orbital primarily exhibits an attractive effect, while the d_xz/d_yz → 2π^* coupling mainly displays a repulsive effect [Figure 2C and D]. Consequently, Ti₂CO₂ shows anomalous d-band adsorption characteristics for ^*N, ^*NH, ^*NH₂, and ^*NH₃ under strain^[57]. A similar phenomenon was also observed during stretching or compressing Pt-V_F-Ti₂CF₂, resulting in a paradoxical situation where the longer Pt–O bond length in ^*OH exhibited a weaker adsorption energy^[58].

Figure 2. (A) Wannier functions of d orbitals and π^* orbital coupling in O₂ molecules with various adsorption configurations; (B) Correlation between adsorption strength and reaction barrier, with variation of -ICOHP for M–O_α and O_α–O_β bonds as a function of $$ d_{x^2-y^2} $$. Copyright 2024, American Chemical Society, Reproduced with permission^[56]; (C) Mechanism of “donation and backdonation” in N₂ adsorption on titanium: interaction between $$ d_{z^2} $$ orbital of Ti and 3σ orbital of N₂, and between d_XZ/d_YZ orbitals of Ti and 2π^* orbitals of N₂; (D) Coupling interactions between adsorbate states and both the average d-band center and spin-polarized d-band center: the complexity arising from d-orbital splitting and its influence on various adsorption behaviors. Copyright 2023, John Wiley and Sons, Reproduced with permission^[57]; (E) Illustration of ORR cycle: a comparison between experimental ORR mechanism and explored five-coordinate model, with the five-coordinate model predicting a closer starting potential to experimental results. Copyright 2022, American Chemical Society, Reproduced with permission^[59]; (F) Illustrative diagram of TM considered in this study, with distinct colors representing corresponding ranges of ORR/OER overpotentials (η); (g) Potential overvoltages for ORR and OER across all TM-N_x-C catalysts; Volcano plots for ORR (H) and OER (I) on the TM-N_x-C system. Copyright 2021, Springer Nature, Reproduced with permission^[60]. ORR: Oxygen reduction reaction; TM: transition metals; OER: oxygen evolution reaction.

Due to incomplete mechanistic understanding of the ORR on Fe-N-C material systems, Hutchison et al. have reported a fifth-coordinate structure for H₂O formation during ORR on Fe-N₄-C, based on a combined study employing DFT and AIMD^[59]. Under potentials relevant to the ORR, OH is converted to H₂O. The results indicated that the Fe(III/II) oxidation-reduction potentials and the ORR onset potentials closely resemble experimental findings. Reliable predictions of ORR onset potential and Fe(III/II) oxidation-reduction potentials are achieved when Fe^III-OH converts to Fe^II, and the desorption of H₂O necessitates axial co-adsorption of H₂O onto the iron center. Considering that the five-coordinate model spontaneously forms and exhibits ORR chemistry consistent with experimental measurements in AIMD simulations, it serves as the fundamental structure for simulating the chemistry of Fe-N-C active sites in experimentally relevant systems [Figure 2E].

Additionally, Xiao et al. employed a combined approach of DFT and AIMD to systematically investigate the catalytic performance of nitrogen-coordinated TM carbon materials (TM-N_x-C) in the ORR and the oxygen evolution reaction (OER)^[60]. They optimized the geometric models of TM-N₃-C and TM-N₄-C using DFT and calculated the formation and binding energies for multiple TM atoms. Ultimately, they selected seven chemically stable metals (Mn, Fe, Co, Ni, Cu, Pd, Pt) for detailed analysis [Figure 2F]. The results from DFT calculations indicated that Ni-N₃-C exhibited the best catalytic performance in the ORR/OER, featuring optimal adsorption energy and the lowest overpotential [Figure 2G-I].

To further validate the reliability of these models, AIMD was used to assess the thermodynamic stability of Ni-N₃-C and Pd-N₃-C at different temperatures. The results shown in Figure 3A and B indicated that Ni-N₃-C remains stable at high temperatures, while Pd-N₃-C is stable only below 400 K, further confirming the potential of Ni-N₃-C as an efficient bifunctional oxygen catalyst.

Figure 3. (A and B) Energy and temperature variations of two catalysts over 5 ps at 400 K vs. the AIMD simulation time. Copyright 2021, Springer Nature, Reproduced with permission^[60]; (C) Free energy diagram of HER over TM@GY SACs. Copyright 2021, Elsevier, Reproduced with permission^[62]; (D) Binding energies of various dopant atoms (Zn, Ni, Mn, Fe, Cu, and Co); (E) DOS for Cu-graphdiyne; (F) Gibbs free energy changes of Co-graphdiyne along the CO₂RR pathway. Copyright 2021, American Chemical Society, Reproduced with permission^[63]; (g) γ-Al₂O₃ (110) and γ-Al₂O₃ (100) crystal surface structures; (H) Surface free energy of γ-Al₂O₃ (110) and (100) crystal planes at different temperatures; (I) Stability test of Ag/(Al-900) sample; (J) Time dependence of bond lengths (Å) during the AIMD simulation (10,000 fs); (K) Snapshot of single Ag atom on the γ-Al₂O₃ (100) surface at different simulation times. Copyright 2024, Springer Nature, Reproduced with permission^[64]. AIMD: Ab initio molecular dynamics; HER: hydrogen evolution reaction; SACs: single-atom catalysts; DOS: density of states.

The aforementioned studies provide strong evidence that DFT and AIMD play crucial roles in the field of catalysis. They serve an irreplaceable function in advancing catalytic theory.

High-throughput computational acquisition data

Compared to traditional experimental exploration methods, DFT significantly enhances the efficiency of catalyst screening but is limited to single calculations. In the bulk screening of efficient catalysts, substantial human effort is still required to organize and submit computational tasks, often leading to resource idling and waste. Therefore, reducing labor input and simplifying repetitive operations has become a critical issue.

High-throughput computing effectively addresses this challenge. By enabling batch submissions of computational tasks, HTS accelerates the screening process as an efficient method for data collection and processing. When combined with DFT and AIMD results, HTS can systematically gather structural and performance data of various SACs, creating a comprehensive database^[78]. Additionally, HTS allows for the rapid selection of samples with target structures from large databases, systematically evaluating catalyst performance. This method can identify and eliminate inefficient catalysts in the early stages, expediting the design and discovery of efficient SACs. By narrowing the experimental scope, HTS not only saves time and costs but also significantly improves the efficiency of identifying effective catalysts.

Comparative studies on the catalytic performance of carbon-based SACs are relatively scarce. To efficiently identify catalysts for the ORR from a multitude of candidates, HTS is a commonly used and effective approach^[79]. Researchers established a database of 48 candidate SACs composed of six TM elements and eight carbon-based supports. Using DFT, they conducted HTS on 180 SACs formed from these eight carbon supports and 3d, 4d, and 5d TM elements^[80]. Using adsorption free energy of OH^* ($$ \Delta G_{OH^{\ast}} $$) as a screening criterion for ORR catalytic activity, accurately predicting the overpotentials on different carbon-based supports and establishing a volcano relationship between $$ \Delta G_{OH^{\ast}} $$ and overpotential [Figure 4A and B]. Through systematic rapid screening, a series of carbon-based SACs with excellent ORR catalytic activity and stability were successfully identified. This study not only elucidated the influence of metal-support interactions on ORR performance but also provided strong theoretical guidance for the design and selection of carbon-supported SACs by saving experimental time and costs.

Figure 4. (A) Summary diagram of ORR overpotential for TM atom-doped different carbon-based carriers; (B) The volcano plot of $$ \Delta G_{OH^{\ast}} $$. The blue regions indicate strong binding energy of OH^*; red dashed lines represent overpotential of Pt (111). Copyright 2021, Springer Nature, Reproduced with permission^[80]; (C and D) The first step of NRR screening results is to evaluate the performance of SACs supported on B-1 for N₂ fixation and conversion to N₂H; (E) The second step screening results of SAC supported on B-1 for NRR based on its NH₂-to-NH₃ formation and desorption performance; (F) Adsorption free energy diagram of H and N₂ on ten selected candidates, with white regions indicating NRR selectivity. Copyright 2022, Elsevier, Reproduced with permission^[81]; (G) The structure of TM@T-C₂N; (H) 3D screening diagram illustrating three types of free energies; (I) Schematic representation of a screening framework for identifying potential NO₃RR catalysts. Copyright 2024, Elsevier, Reproduced with permission^[82]. ORR: Oxygen reduction reaction; NRR: nitrogen reduction reaction; SACs: single-atom catalysts.

HTS can provide a valuable reference framework for screening of other multi-step reactions. Yue et al. investigated the nitrogen reduction reaction (NRR) performance of four surface termination structures WB₂(001) using a two-step HTS methodology [Figure 4C-F]. In the first step, they calculated the adsorption energies of N₂ and H₂, along with the free energies of the reaction pathways, to identify SACs with potential NRR activity. In the second step, they further calculated the free energies for NH₃ synthesis and desorption to select SACs with excellent NRR performance^[81]. Through this two-step screening process, they identified four notably effective SACs: V@(B-1@3), Mn@(B-2@3), Cr@(W-1@1), and Cr@(W-2@1).

Using HTS combined with first-principles calculations, researchers investigated the application of TM-tetragonal carbon nitride (TM@T-C₂N) catalysts [Figure 4G] in the electrochemical nitrate reduction reaction (NO₃RR)^[82]. They proposed a five-step screening criterion [Figure 4H], outlining the elimination sequence and stages for potential NO₃RR candidates based on different reaction phases.

The Gibbs free energy descriptors from the first three steps were used to construct a three-dimensional (3D) screening map, as shown in Figure 4I. The pink, blue, and yellow cross-sections visualize the screening criteria: $$ \Delta G_{{ }^{\ast}NO_3} $$ < 0, $$ \Delta G_{{ }^{\ast}NH_2\to { }^{\ast}NH_3} $$ < 0.7, and ΔG_6-min< 0.70 eV, respectively. This process retained six candidate materials (Ti-, V-, Cr-, Mn-, Zr-, and Tc@T-C₂N) for further consideration. These criteria facilitated the rapid exclusion of unsuitable candidates, narrowing the focus of the study and allowing for the efficient pre-screening of a large number of candidates. Ultimately, V@T-C₂N and Cr@T-C₂N were selected as potential high-performance NO₃RR catalysts.

In a word, DFT and AIMD provide an atomic-level perspective for understanding catalyst performance, while HTS enables rapid evaluation of numerous candidates based on specific criteria, effectively reducing research time and scope. The integration of these methods offers robust support for the efficient development of SACs. It remains noteworthy that discrepancies may exist between the models employed in HTS and actual material structures, leading to deviations in the prediction results. Materials with high performance identified through screening may encounter difficulties during experimental preparation, and the screening outcomes necessitate extensive experimental validation.

Development stage: feature importance analysis using regression models

HTS generates a wealth of computational data, providing a foundational basis for subsequent analysis and model building^[83-85]. However, it does not directly assess the importance of factors influencing catalyst activity. For example, in Pt-doped Janus-MXenes, the binding energy between Pt atoms and the substrate, work function, and the number of electrons gained by Pt atoms are closely related to catalyst activity^[86]; however, the ranking of importance of different features remains unclear. Similar reports indicate that there is a strong linear relationship between the d-band center of metal atoms, the number of electrons transferred, and catalytic activity^[34]. However, this relationship cannot be quantified in percentage terms to express the importance of the d-band center.

ML, particularly regression models, can predict material performance from existing data^[87,88]. Common regression models, such as random forests, support vector machines, and decision trees, can handle multidimensional data and analyze feature contributions to catalytic performance. ML can predict new catalyst performance based on feature engineering, accelerate screening process of SACs and provide a theoretical basis for designing efficient catalysts, making research more targeted. Furthermore, the importance analysis of features deepens the understanding of how different characteristics influence catalytic performance, expediting the identification of high-efficiency catalysts.

Application of regression models to the performance analysis of SACs

Among various ML models, regression models require fewer data points and can perform feature importance analysis^[89]. Therefore, they are more suitable for integration with DFT calculations and HTS of SACs.

In this context, Chen et al. introduced a method for rapidly screening CO₂ reduction electrocatalysts based on simple features and ML models^[90]. The researchers constructed a database consisting of 1,060 metal-nonmetal CO-doped graphene structures, and optimized the feature set through Pearson correlation heatmaps [Figure 5A] and feature importance ranking [Figure 5B]. This led to the identification of an optimal feature set containing eight features. Various ML algorithms were tested, including K-nearest neighbors (KNN), random forest regression (RFR), support vector regression (SVR), gradient boosting regression (GBR), extreme GBR (XGBR), and a kind of composited algorithms produced by tree-based pipeline optimization tool (TPOT) [Figure 5C-H], with cross-validation used to evaluate the prediction performance of different algorithms. Among all the algorithms tested, the XGBR exhibited superior predictive performance, characterized by a higher coefficient of determination (R²) value and a lower root mean square error (RMSE) value. When compared to the composite algorithms generated by TPOT, XGBR boasts a simpler structure, higher controllability, and greater ease in hyperparameter tuning and model optimization. The predictive model established based on the XGBR model successfully predicts changes in CO adsorption free energy (ΔG_CO), thereby enabling the evaluation of the catalytic activity. Moreover, the XGBR model showed excellent generalization ability. To assess the impact of the HER on CO₂ reduction, another XGBR model was developed specifically to predict the HER catalytic activity of the 1,060 materials. By combining the predictions from both models, 94 potential CO₂ reduction electrocatalysts were successfully screened.

Figure 5. (A) Heatmap of Pearson correlation coefficient matrix for the ΔG_CO- predicted optimal feature set; (B) Ranking of feature importance within the optimal feature set; (C-H) Predictive performance of various models trained using different ML methods. Copyright 2020, American Chemical Society, Reproduced with permission^[90]; (I and J) Feature importance of the top ten significant features predicted by GBR and XGBR models; (K) Utilizing SHAP analysis to consider the overall impact of different features on model prediction; (L) Predicting reaction free energy via the GBR model: excellent agreement between predicted values and DFT calculations. Copyright 2024, Elsevier, Reproduced with permission^[91]. ML: Machine learning; GBR: gradient boosting regression; XGBR: extreme gradient boosting regression; SHAP: Shapley Additive Explanations; DFT: density functional theory.

Additionally, Zhang et al. employed interpretable ML models to directly predict the Gibbs free energy for evaluating the electrocatalytic nitrogen reduction activity of SACs^[91]. It is noteworthy that all the features used in the model were not derived from DFT calculations. Instead, a dataset of 90 graphene-based TM SACs was collected from available literature, which included the catalyst’s structure and reaction free energies, along with 41 basic features extracted from the periodic table, such as atomic number, atomic mass, covalent radius, and d-electron count. Pearson correlation heatmaps were used to eliminate highly correlated features, and independent features were selected for model training. Based on the independent feature set database, various ML models, including GBR, XGBR, and RFR, were successfully trained. Ultimately, GBR was selected as the optimal ML model for predicting Gibbs free energy of NRR. It demonstrated the highest prediction accuracy during both training and testing phases with an R² score exceeding 0.97 and RMSE less than 0.1 eV. These performance metrics significantly outperform other models, such as XGBR and RFR. In terms of interpretability, feature importance analysis and Shapley Additive Explanations (SHAP) can uncover the working mechanisms of GBR model, providing insights into the specific influence of individual features on prediction outcomes and guiding the SAC design. Figure 5I and J show the feature importance analysis, emphasizing feature importance within the context of the model structure and training process. Figure 5K presents the SHAP analysis, which considers the overall impact of different features on model predictions, accounting for feature interactions and correlations. GBR possesses exceptional feature capture capabilities, enabling effective identification of key features related to the active center and coordination environment, such as the radius of TM (r_TM), average radius of TM (r_TM-ave), number of d-electrons (N_d), and heat of formation (H_of). It was evident that the covalent radius of active center atom was consistently one of the most important features influencing NRR activity.

The analysis results based on the GBR can guide the improvement of SACs through considerations of coordination types, d-electron count, and covalent radius. For instance, selecting SACs with pyrrole-type coordination (flag = 0) can significantly reduce Gibbs free energy for NH₃ desorption, thereby enhancing NRR activity. Furthermore, by optimizing d-electron count and the difference in covalent radius, it is possible to enhance N₂ activation, suppress HER, and improve selectivity of NRR. Figure 5L demonstrates the predicted reaction free energies obtained through GBR align well with those calculated by DFT, validating the reliability of the model.

The studies by Chen et al. and Zhang et al. highlight the significant role of regression models in predicting catalyst performance^[90,91]. Regression models can rapidly screen potential materials, exhibiting significantly higher efficiency compared to HTS. This approach drastically reduces research costs by eliminating the need for extensive DFT calculations. Feature importance assessment can quantify the impact of different features on the catalytic performance of SACs, helping identify and focus on the most informative features. These results not only accelerate the discovery and screening process of SACs but also help understand the internal mechanisms of ML models. Furthermore, the trained ML models exhibit excellent transferability and robustness, making them powerful tools for future catalyst research and development.

Feature importance analysis accelerates the screening process of SACs and provides a theoretical foundation for designing efficient catalysts, making the research process more targeted^[92-94]. In a recent study, Pritom et al. combined DFT and ML methods to investigate the influence of porphyrin-supported SACs on the charging behavior of non-aqueous magnesium-carbon dioxide (Mg-CO₂) batteries^[95]. The study reveals the high efficiency of SACs in both the binding of MgCO₃ and promoting its decomposition. By employing a ML model, the adsorption energy of MgCO₃ on SACs was successfully predicted, and the key factors influencing the adsorption energy were analyzed. The results indicate that the ionization potential of the TM is a crucial parameter for selecting SACs.

In the process of constructing the ML model, the authors selected electronic structure features closely related to the adsorption performance of SACs as inputs. These features include the distance between the TM and N/S, the charge of the TM, the N_d, the average charge of nitrogen, electronegativity, ionization energy, and electron affinity. Correlation analysis using the Pearson correlation coefficient was performed to select independent features and identify redundant ones. For example, it was found that the TM charge was significantly correlated with other features, so it was removed from the model. Additionally, the authors used violin plots [Figure 6A and B] to show how the adsorption energy of MgCO₃ exhibited different distributions at various levels under different environments, clearly revealing the heterogeneity of the data distribution and the differences across configurations, providing an intuitive basis for feature selection.

Figure 6. (A and B) Violin plots illustrating the adsorption energies of MgCO₃ across various environments; (C and D) Graphic representation of feature importance utilizing MDI from GBR model and permutation importance technique from ANN model. Copyright 2024, Royal Society of Chemistry, Reproduced with permission^[95]; (E) Representation of predictions generated by a training model using a 3D matrix composed of the three most influential descriptors, with four distinct regions classified by k values, highlighting a matrix of calcination and pyrolysis temperatures and enlarging the area with high k values; (F) Violin plot of SHAP values for the XGBR model (left) and summary plot of SHAP analysis (right). Copyright 2023, American Chemical Society, Reproduced with permission^[41]; (G) Illustrating the cruciality of six features through gain value and SHAP value; (H) SHAP summary plot for ML models; (I and J) Violin plots of SHAP values for N_V and SHAP dependence for N_n, with blue indicating eligible catalysts and red indicating ineligible ones. Copyright 2023, John Wiley and Sons, Reproduced with permission^[100]. MDI: Mean decrease in impurity; GBR: gradient boosting regression; ANN: artificial neural network; SHAP: Shapley Additive Explanations; XGBR: extreme gradient boosting regression; ML: machine learning.

To predict the adsorption energy, the authors selected two models, GBR and artificial neural networks (ANNs), due to their advantages in handling nonlinear relationships and complex data. Prior to model training, the data was standardized to ensure comparability between different features, and K-fold cross-validation was employed to assess the model’s generalization ability and avoid overfitting. By training the GBR and ANN models with DFT-calculated results and adjusting the model parameters and structures, the authors achieved high prediction accuracy.

To gain deeper insights into model predictions, the authors employed multiple methods such as SHAP, permutation importance, and mean decrease in impurity (MDI) [Figure 6C and D] to analyze feature importance. These methods evaluated the influence of each feature on the model’s predictions from different perspectives and revealed that ionization energy and the N_d were key features influencing the MgCO₃ adsorption energy. It suggests that future improvements in the catalytic performance of SACs can be achieved by optimizing these key features.

For the optimization of multi-step chemical transformations, a ML framework has been developed to guide catalyst design by analyzing key steps in the multi-step process to enhance reaction efficiency^[96-99]. In another study, Fu et al. reported the use of ML algorithms to accelerate the design process of highly efficient Fenton-like SACs^[41]. The XGBR prediction model built using ML algorithms accurately predicted the degradation rate (k-value) of SACs for phenol. The SHAP explanation method quantified the impact of various parameters on the model’s predictions, as shown in Figure 6E and F, revealing that Fe loading, carbonization temperature, and carbonization heating rate are key factors influencing the k-value. Through ML-guided optimization, they identified efficient SACs dominated by Fe-N₅ sites, exhibiting excellent Fenton-like activity (k = 0.158 min^-1). This work provides an example of ML-assisted optimization of single-atom coordination environments and demonstrates its feasibility in accelerating the development of high-performance catalysts, thereby helping researchers gain a deeper understanding of the structure-performance relationship of SACs.

Materials with high performance identified through screening may encounter difficulties during experimental preparation, and the screening outcomes necessitate extensive experimental validation. Sun et al. proposed a strategy based on interpretable ML and HTS to understand and predict the performance of NRR catalysts^[100]. Through DFT calculations, a catalyst space containing 168 types of carbon-supported SACs was constructed. A four-step screening strategy was employed to identify 33 promising candidate catalysts. Using the interpretable ML model XGBR, the key features influencing NRR performance were analyzed. The feature importance, SHAP value distribution, and dependency plots [Figure 6G-J] revealed that the valence electron count (N_V) and nitrogen substitution number (N_n) of the metal atom are the two key factors affecting catalytic performance. Further analysis showed that active centers with lower N_V and moderate N_n could better balance charge distribution, thereby enhancing NRR catalytic activity. Ultimately, six catalysts with excellent NRR performance were selected, and reaction pathways and rate-determining steps were analyzed.

To sum up, ML models are crucial for SAC studies, enabling precise prediction and optimization of catalyst performance, thus enhancing design efficiency and understanding of catalytic mechanisms. AI methods rely heavily on data quality and quantity, which are costly and hard to obtain. Many ML models, especially deep learning, lack interpretability, making it difficult to understand the physical mechanisms behind predictions. Additionally, their generalization ability may be limited when faced with new, unseen data. As ML advances, continuous optimization of algorithms and models will improve accuracy and reliability, driving the progress of catalytic science.

Growth stage: using NN to analyze the structural characteristics of SACs and screen high-performance catalysts

By using ML regression models, researchers have successfully identified key features that influence catalyst activity and conducted feature importance analysis^[36,101-103]. A critical challenge now is leveraging these important features to screen and predict the catalytic performance of new structures, which is key to simplifying the screening process and reducing costs.

NNs [such as deep neural networks (DNNs), convolutional neural networks (CNNs), and graph neural networks (GNNs)] are particularly well-suited for this task due to their layered structure, with each layer containing multiple neurons. These networks can process input data through activation functions, capturing complex nonlinear relationships. By automatically learning high-dimensional features, NNs can better describe and predict the properties of materials^[104], making them ideal for handling complex models and large-scale datasets. Consequently, NNs hold great promise for screening high-performance catalysts from complex structures.

DNNs can significantly reduce computational time, quickly eliminating large numbers of ineffective catalysts, and saving computational resources^[105-107]. At the same time, these models help reveal the relationship between structural features and reaction mechanisms^[39,108]. Zafari et al. utilized coulomb matrices and principal component analysis (PCA) to reduce the dimensionality of geometric structure features of SACs, which helped extract relevant features and reduce model complexity^[109]. The optimized structural information, consisting of seven features, was used as the input for the DNN model with an input layer, two hidden layers, and an output layer [Figure 7A]. The DNN model predicts N₂ adsorption energy and hydrogenation free energy, successfully screening three NRR electrocatalysts-CrB₃C₁, TcB₃C₁, and HfB₁C₂ with low overpotential and high selectivity.

Figure 7. (A) An illustrative diagram of the ANN architecture (featuring 10 neurons per hidden layer), utilizing optimized SAC geometric models as input data, with each structural geometry possessing seven distinct features. Copyright 2020, Royal Society of Chemistry, Reproduced with permission^[109]; (B) Classification of SACs into three categories using PCA and K-means clustering of XANES data (with different colors representing distinct clusters); (C) Comparison of experimental XANES spectra (in black) with theoretical XANES spectra reconstructed from descriptor values predicted by NN; (D and E) The prediction accuracy of NN models on the test dataset. Copyright 2022, Royal Society of Chemistry, Reproduced with permission^[110]; (F) Volcano plot and CNN-based catalyst performance analysis pipeline: integrative use of volcano plots for predictive assessment of existing catalysts, with eDOS as input for predicting and tuning adsorption energies, and extraction of chemical information from CNN model; (G) Prediction of adsorption energies for intermediates in the CO₂RR process using a CNN model; (H and I) Limiting potential volcano plot and periodic table. Copyright 2024. This publication is licensed under CC-BY-SA 4.0^[112]; (J) AC-STEM image of Pt1/NC with ML-detected overlapping atoms highlighted in yellow circles. Magnified view emphasizing potential elements for detection and quantification (indicated by blue circles); (K) The representative prediction maps generated by CNN for elements in (J), with interatomic distance analysis in the left image and overlapping features addressed using Gaussian Mixture Models assignment in the two images on the right; (L) Inference of corresponding atomic position assignments by CNN models and prediction of atomic chemical properties through VAE latent space clustering; (M) Comparison of performance metrics for ML models versus manual tasks executed by domain experts: ML model detection on test images accomplished in minutes versus hours required for human expert tasks. Copyright 2023, John Wiley and Sons, Reproduced with permission^[116]. ANN: Artificial neural network; SAC: single-atom catalyst; PCA: principal component analysis; XANES: X-ray absorption near-edge structure; NN: neural network; CNN: convolutional neural network; eDOS: electronic density of states; AC-STEM: aberration-corrected scanning transmission electron microscopy; ML: machine learning; VAE: variational autoencoders.

X-ray absorption near-edge structure (XANES) analysis is a powerful technique for probing the structural changes in SACs. However, traditional XANES analysis struggles to handle structural heterogeneity, making it challenging to accurately identify the number of SAC species. In the study by Xiang et al. on photocatalytic CO₂ reduction, the NN model could extract structural information from XANES data, enabling a more accurate and efficient identification of SAC species and structural variations during reaction processes^[110].

Using methods such as PCA, K-means clustering, and NNs, the XANES data was successfully analyzed to obtain quantitative structural information about the local atomic environment of SACs. This approach allowed the classification of SAC species into three categories [Figure 7B], identifying their number and providing a foundation for subsequent structural analysis. By employing the NN-XANES method, the local geometry of Co-cyclaml-CO was refined. A large set of theoretical XANES data was trained to establish a mapping relationship between the XANES features and structural descriptors. The trained NN model showed consistency with experimental data [Figure 7C-E], and it could predict structural descriptors from experimental XANES data, such as bond lengths and bond angles. Additionally, it provided more detailed structural information, allowing for the distinction of Co-O and Co-N contributions.

CNNs are a type of feedforward NN composed of convolutional layers, pooling layers, and fully connected layers^[111]. With features such as local connections, weight sharing, and pooling, CNNs excel in processing images or structural data of materials. Unlike traditional methods such as the d-band center model, CNNs automatically extract features from electronic DOS (eDOS) without manual intervention, establishing complex relationships with adsorption energy. Yang et al. proposed a workflow that combines CNNs with volcano plots [Figure 7F] to screen and predict two-dimensional SACs in CO₂RR^[112]. By establishing correlation plots between intermediate adsorption energies and various descriptors, a CNN model was used, with 2D eDOS as input, to predict adsorption energies and understand the impact of electronic structure perturbations. The CNN model predicted the adsorption energies of nine intermediates in CO₂RR with an average mean absolute error (MAE) of 0.06 eV [Figure 7G], demonstrating high prediction accuracy compared to DFT. It also exhibited strong generalization ability in handling species containing oxygen, hydrogen, carbon atoms, and different substrates. In addition, a hybrid descriptor combining C-type and O-type CO₂RR intermediates was introduced to construct optimized volcano plots and periodic tables [Figure 7H and I], providing a visual approach to efficiently screen for CH₄ reduction catalysts.

CNNs also play a crucial role in the structural characterization of SACs, enabling rapid, accurate, and automated detection of metal centers^[113-115]. Due to the current lack of research focused on metal centers, it is challenging to design atomically precise structural materials. However, the use of CNNs enables rapid and standardized detection of metal centers^[116]. CNNs can identify pixel patterns in aberration-corrected scanning transmission electron microscopy (AC-STEM) images, distinguishing between metal atoms and background pixels accurately. Threshold segmentation and bounding box recognition techniques allow for thresholding of the probability map output by CNNs and the use of bounding boxes to identify the coordinates of metal atoms [Figure 7J and K]. CNNs and Gaussian mixture models can perform chemical specificity analysis on multi-metal ultra-high-density (UHD)-SACs [Figure 7L], distinguishing metal centers of different chemical types and quantifying mixing degrees between metal centers. Compared to manual methods, CNNs demonstrate higher accuracy and repeatability, as shown in Figure 7M, greatly improving detection efficiency.

GNNs further expand the application of NNs and are tailored for processing graph-structured data, particularly useful for handling molecular or crystal structures^[117]. By representing material structure as graph networks with nodes (atoms) and edges (bonds), GNNs can efficiently extract both local and global structural information, propagating information between adjacent nodes and learning complex structural relationships^[118]. For SACs, GNNs can model the interactions between metal atoms and supports, predicting catalytic performance based on these interactions.

In material design, surface structural changes at the nanoscale are especially important. Small molecule adsorption energy is a key indicator of catalyst activity, but linear scaling relations limit performance improvement. Surface strain can break these scaling relationships. Surface strain engineering involves a high-dimensional search space, and comprehensive DFT screening is impractical. GNNs can efficiently handle high-dimensional data by learning nonlinear functions and generalizing from relatively small training datasets, allowing for efficient exploration of the strain space. Using GNNs to predict the adsorption energy response of catalyst/adsorbate systems under surface strain patterns, Price et al. proposed a GNN model [Figure 8A], effectively bridging the gap between experimental and theoretical results^[119]. The normalized confusion matrix of the GNN + strain model for experimental data [Figure 8B] successfully predicts the strain patterns in 85% of unseen test data, outperforming linear models. The model also predicted strain responses in ammonia synthesis reaction intermediates [Figure 8C], revealing the role of compressive strain in breaking linear scaling relations. This study provides a new approach for identifying strain patterns that can break the adsorption energy scaling relationships. By generating phase diagrams of adsorption energy versus strain [Figure 8D and E], it offers an intuitive method for strain engineering, guiding catalyst design and improving performance.

Figure 8. (A) The architecture of GNNs for classification and regression tasks; (B) The normalized confusion matrix for test data. Each row corresponds to different true classes, and each column corresponds to predicted classes; the diagonal represents the percentage of correct predictions for each class; (C) Prediction diagram of strain response for single-molecule NH₃ synthesis on Cu₄S₂ (110) surface by regressor; (D and E) Verification and comparison of strain phase diagrams for HfCu₃(100) surface adsorption with ^*N and ^*NO₂ using DFT. Copyright 2022, The American Association for the Advancement of Science, Reproduced with permission^[119]; (F) Diagram of the GNN architecture applied in this study; (G) The workflow diagram of combining ML with DFT screening; (H) The pairing diagram of CGCNN model, DFT-calculated $$ \Delta G_{H^{\ast}} $$, and ML-predicted $$ \Delta G_{H^{\ast}} $$. Copyright 2023, American Chemical Society, Reproduced with permission^[120]. GNNs: Graph neural networks; DFT: density functional theory; ML: machine learning; CGCNN: crystal graph convolutional neural network.

It is clear that both CNN and GNN models have their own strengths. The crystal graph convolutional neural network (CGCNN) combines the advantages of both CNNs and GNNs, enabling it to learn material properties from atomic connections within crystals and providing highly accurate predictions. CNNs excel at processing image data and can be used to identify the crystal structures of materials. In contrast, GNNs are well-suited for handling graph-structured data to capture atomic interactions and chemical bonding information.

Figure 8F indicates that the CGCNN accelerates high-performance dual-atom catalyst (DACs) screening by learning the structure-activity relationships of existing DACs, enabling the prediction of $$ \Delta G_{H^{\ast}} $$ and facilitating the selection of graphene-based DACs with ideal conductivity and stability for HER, thereby reducing the computational cost of DFT by half^[120]. By combining CGCNN with DFT [Figure 8G], the model screened 435 dual-atom combinations of nitrogen-doped graphene (N₆Gr) and identified two high-performance HER catalysts (AuCo@N₆Gr and NiNi@N₆Gr). DFT confirmed these two catalysts exhibited excellent reaction kinetics, and an analysis of electronic states, $$ \Delta G_{H^{\ast}} $$, and metal-hydrogen interaction strength revealed the crucial role of metal synergy in HER.

CGCNN avoids the need for feature engineering by directly learning the chemical structure from the material’s geometric configuration, which simplifies the screening process and saves computational resources. The $$ \Delta G_{H^{\ast}} $$ predicted by the CGCNN model shows a high agreement with the DFT results, with a MAE and RMSE of 0.22 and 0.27 eV, respectively [Figure 8H], demonstrating the model’s high accuracy. The CGCNN is effectively applied to the screening of DACs, contributing to the acceleration of catalyst discovery.

Overall, both CNNs and GNNs offer significant advantages in analyzing complex catalyst structures. By learning high-dimensional features automatically, they improve the description and prediction of material properties, advancing materials science. These methods help researchers identify promising high-performance catalysts from numerous SAC combinations, significantly reducing experimental efforts. Future work should address the high computational costs and complex training processes of NNs for large-scale material structures. Additionally, attention is needed to prevent over-smoothing in deep, multi-layer networks, as it can reduce model performance.

Maturity stage: design the SAC structure and predict its performance with a generation model

In the mature phase of integration between AI and materials science, generative network models have demonstrated immense potential in catalytic structure design. In this reverse design process, users can define catalytic properties of SACs and generate model structures with precisely defined attributes. For instance, generative models such as GANs and variational autoencoders (VAE) can learn low-dimensional representations from training data and continuously alter parameters^[121,122]. These models coupled with computational methods such as DFT to validate whether the generated structures meet required performance criteria, thereby optimizing the design. In specific applications, generative models extract synthesis steps and catalytic properties from literature, employ active learning to explore the chemical space of specific catalysts, and leverage models such as GANs and VAE to generate hypothetical alloys and ligands^[123].

Generative models offer an efficient, diverse, and interpretable approach to swiftly generating and evaluating a variety of catalyst structures, thus accelerating the discovery and design process^[124,125]. GANs can help automate the improvement of catalyst materials by generating new catalyst surfaces with higher activity. Ishikawa proposed a novel approach that combines computational chemistry and ML to “extrapolate” new catalytic surfaces^[126]. DFT is used to calculate the energy of basic reactions on a given set of catalyst materials, and then the results are input into the GANs.

Using a GAN trained on a DFT dataset, researchers have successfully generated more complex and diverse catalyst structures, expanding the possibilities for catalyst design [Figure 9A]. These newly generated surface structures, not included in the initial dataset, exhibit higher turnover frequency (TOF) values for the ammonia synthesis reaction. Through iterative training [Figure 9B], the model continuously learns patterns and trends from the DFT dataset, applying them to create novel catalyst surface structures. This approach leverages GANs in combination with DFT calculations to “extrapolate” catalysts with enhanced catalytic activity, optimizing key factors such as reaction energy and activation energy.

Figure 9. (A) Flow chart of the DFT-GAN program structure: training and evaluation phases; (B) Discriminator and generator losses in DFT-GAN, with each iteration comprising 2,000 epochs; (C) The distribution maps of Ru and Rh atoms on the initial surface (iter = 0) and GAN-generated surfaces (iter = 1-5); (D) The TOF of NH₃ formation on Rh-Ru alloy surface, with TOF values from different DFT-GAN iterations (iter = 1-5) encoded in distinct colors; (E) Box plot and violin plot of TOF values for iter = 0-5. Left-side points on violin represent raw TOF values for each iteration. Copyright 2022, Springer Nature, Reproduced with permission^[126]; (F) Application potential of DL techniques in image-based catalyst screening, with recognizable image types including chemical images, morphological images, and catalytic images; (G) The workflow diagram of ML and DL for discovering HER electrocatalysts. Copyright 2021, American Chemical Society, Reproduced with permission^[127]; (H) Flow chart of the VAE network in AGoRaS: decompression of chemical database information into a high-dimensional latent space; (I) The flow chart of data collection, training, and validation steps employed by AGoRaS for generating synthetic data; (J) t-SNE visualizations of training and generated datasets. Upper panel: a sample of 7,000 equations from the training dataset, alongside 7,000 randomly selected equations from the generated dataset. Lower panel representation of 70,000 equations extracted from the generated dataset. Copyright 2022, Springer Nature, Reproduced with permission^[128]. DFT: Density functional theory; GAN: generative adversarial network; TOF: turnover frequency; DL: deep learning; ML: machine learning; HER: hydrogen evolution reaction; VAE: variational autoencoders; t-SNE: t-distributed stochastic neighbor embedding.

In this method, DFT calculations are used to compute the energy (ΔE) of elementary reactions for all surfaces present in the initial dataset. TOF values for ammonia synthesis are then obtained from ΔE values, and metal surfaces are labeled according to their TOF. A GAN, composed of a discriminator and generator, is trained on this DFT dataset, enriched with TOF values and metal surface information.

The generator of the GAN creates samples not contained in the current dataset. In this case, a conditional GAN is employed to generate surfaces with higher TOF values. Figure 9C demonstrates the GAN to learn and recognize key factors affecting catalyst activity, such as step sites adjacent to adsorbing atoms and to apply this knowledge to produce new catalyst surface structures with enhanced activity. DFT calculations are then performed on the newly generated samples, with results integrated back into the dataset. Figure 9D and E illustrate the iterative process, starting with 100 random steps and alloy surfaces created through atomic substitution. After five iterations, a previously unobserved Rh₈Ru₇₆ surface was successfully obtained, achieving a TOF more than ten times higher than the best TOF value in the original dataset.

All in all, samples generated in later iterations typically exhibit higher TOF values, indicating that the iterative DFT-GAN approach effectively enhances NN training within the GAN. Moreover, the generated surfaces tend to show a higher proportion of Ru atoms, aligning with experimental observations. This improvement is attributed to a lower activation energy for the RDS due to the reduced dissociation energy of N₂ and a lower formation energy of NH₃, which decreases NH₂ coverage on the surface and mitigates NH₂ poisoning. These characteristics contribute to higher TOF values in the generated surfaces. This study demonstrates that combining DFT with GAN is a promising strategy for the automated, continuous improvement of catalyst performance. Compared to traditional catalyst design methods, GANs enable the rapid and efficient generation of novel catalyst surfaces with higher catalytic activity, facilitating catalyst material optimization without the need for manual intervention, thus enhancing both the efficiency and accuracy of catalyst design.

Deep learning image-based recognition offers distinct advantages in processing various types of image data, such as chemical, morphological, and catalytic images. Figure 9F highlights the potential of deep learning techniques in image-based catalyst screening. These images can provide valuable information, enabling researchers to rapidly identify and screen efficient TM-based electrocatalysts for HER. By introducing the latest advancements for identifying highly active TM-based HER electrocatalysts, representative studies utilize deep learning NN architectures [Figure 9G] to screen catalysts based on chemical and morphological images. This approach enhances the understanding of the relationship between the intrinsic properties of TM-based materials and their electrocatalytic performance^[127].

This generative design approach transforms AI from merely a tool for predicting material properties to an active driver in the discovery of new materials. Tempke and Musho have introduced an AI model known as AGoRaS, based on a VAE, designed for synthesizing new chemical reaction datasets^[128]. The AGoRaS VAE model is structured with embedding layers, bidirectional long short-term memory (LSTM), and latent space sampling, enabling the generation of an unbiased chemical reaction dataset by encoding reaction data into a latent space [Figure 9H]. By sampling within this latent space, the model generates novel chemical reactions, circumventing the biases typically present in traditional datasets. The model incorporates a rigorous data collection, training, and validation pipeline to ensure the validity and stability of generated reactions [Figure 9I].

Figure 9J illustrates a t-distributed stochastic neighbor embedding (t-SNE) plot that reduces high-dimensional data to two dimensions, showcasing the model’s application on a training dataset of 7,000 reactions. This approach successfully produced over seven million new reactions, including 20,000 molecular species previously absent from the dataset, broadening the predictive capabilities of ML algorithms. Additionally, the AGoRaS model predicts thermodynamic properties for new reactions, such as Gibbs free energy, entropy, and dipole moments. Semi-empirical calculations help validate the stability of the predictions. By enabling targeted reaction searches that include specific molecular species, this model provides new avenues for experimental research. AGoRaS is thus poised to facilitate the generation of novel chemical reactions, guiding experimental efforts while enhancing the robustness and accuracy of AI algorithms in chemical science.

ML, coupled with genetic algorithms and Bayesian optimization, enables the continuous evolution of catalyst structures, optimizing performance. This approach accelerates SAC discovery and facilitates custom designs for clean energy and environmental protection. By leveraging structure-performance databases, GANs generate efficient catalysts with novel atomic arrangements, enhancing activity and stability for specific reactions such as CO₂ reduction and oxygen reduction.

The demand for novel functional materials necessitates effective strategies to expedite material discovery, where crystal structure prediction is foundational^[129,130]. Generative models such as GANs serve as a powerful means to explore hidden regions within chemical space^[122]. Kim et al. proposed a GAN-based approach for crystal structure prediction^[131], structured with a generator, discriminator, and classifier. The generator produces new crystal structures from random noise vectors and encoded composition vectors; the discriminator computes the Wasserstein distance between real and generated data to assess authenticity, and the classifier ensures the generated structures align with specified composition. Applying the GAN model to Mg-Mn-O systems enabled generative high-throughput virtual screening for photoanode properties. The crystal structures generated by the GAN model demonstrated reasonable stability and Heyd, Scuseria, and Ernzerhof (HSE)-calculated band gaps, some of which possess unique configurations compared to existing materials.

Another innovative approach, the CGCNN, has been developed for predicting the properties of crystalline materials^[132]. This method represents crystal structures as graphs and builds CNN upon them [Figure 10A]. CGCNN learns and automatically extracts atomic connectivity features within crystal structures to predict various material properties, such as formation energy, band gap, and elastic properties. By leveraging CGCNN, researchers can estimate energy contributions of atoms within perovskites, uncovering empirical design rules of materials.

Figure 10. (A) CGCNN process diagram. CGCNN converts crystal structures into feature vectors, learning and predicting material properties. Copyright 2018, American Physical Society, Reproduced with permission^[132]; (B) Process diagram for building predictive models using ML. This performance model holds the capability to forecast catalytic-related properties based on computational data and information sourced from material databases; (C) The general steps of catalyst optimization genetic algorithm supported by AI. Copyright 2024, American Chemical Society, Reproduced with permission^[136]; (D) Roadmap for generating NES; (E) Workflow of Bayesian Optimization algorithm. Copyright 2021, OAE Publishing Inc. Reproduced with permission^[137]; (F) Initial state model with limited data points; (G) Advanced stage of optimization, model improved through a larger dataset; (H) Predicted optimal point by Bayesian optimization algorithm, along with experimental data points obtained. Copyright 2023, American Chemical Society, Reproduced with permission^[141]. CGCNN: Crystal graph convolutional neural network; ML: machine learning; AI: artificial intelligence; NES: neural evolutionary structures.

In summary, generative models offer distinct advantages in exploring chemical space, especially in uncovering regions that traditional methods miss. They enable researchers to evaluate the impact of local chemical environments on global properties and perform combinatorial searches for synthesizable materials, providing valuable insights and design rules. This approach narrows the search and accelerates discovery.

GANs can pinpoint local maxima within the material design space, guiding materials toward optimal performance. For SACs, catalytic properties are closely linked to the local environment of metal atoms on support materials, including electronic structure, atomic spacing, and coordination numbers^[133,134]. By simulating these local environments, GANs can generate catalyst models with ideal electronic structures, achieving improved activity and selectivity in catalytic processes. Moreover, the generative and discriminative mechanisms of GANs allow the swift screening of numerous candidate structures, making this approach especially effective for optimizing heterogeneous catalysts, where traditional experimental screening would be prohibitively costly and time-intensive.

On the data-driven front, AI/ML can develop performance surrogate models that generate catalyst structures with specific chemical properties or reaction pathways based on given input conditions^[135]. These models can also predict the properties of unknown substances by training NNs to correlate input parameters with catalytic performance, enabling the identification of new active sites and guiding the optimization of experimental conditions [Figure 10B]. However, the field faces challenges, such as system complexity, data diversity, and target variability. To address these, techniques such as genetic algorithms [Figure 10C] and Bayesian optimization are explored to guide optimization, providing structured approaches to refine catalyst design amidst these complexities^[136].

NNs face limitations, including extensive data and computational needs, and lack of interpretability, necessitating integration with other technologies. High-entropy alloy catalysts are notable for superior catalytic performance; however, traditional trial-and-error methods hinder systematic exploration of structure-performance relationships. By training NNs on small unit structures and using inverse design for larger structures [Figure 10D], high-entropy alloy models can be efficiently generated to analyze structure-performance relationships. NNs predict key catalytic features, such as adsorption energies at active sites, identifying the most influential sites. Combined with Bayesian optimization [Figure 10E], this accelerates discovery of high-performance high-entropy alloy catalysts by automatically identifying optimal compositions and predicting activities^[137].

An innovative AI-driven catalyst optimization workflow, incorporating large language models (LLMs), Bayesian optimization, and active learning loops, can accelerate catalyst optimization^[138-140]. This workflow effectively combines advanced language understanding with robust optimization strategies, translating knowledge extracted from diverse literature into practical parameters for experimentation and optimization. Lai et al. demonstrated the application of AI workflow in the synthesis of ammonia synthesis catalysts; the results indicated that this workflow simplifies the catalyst development process^[141]. Figure 10F and G illustrate that Bayesian optimization leverages the information extracted by LLMs to approximate the unknown complex function mapping catalyst synthesis parameters to performance indicators through the construction of probabilistic models, which are continuously updated through active learning loops, ultimately converging to the global optimal solution [Figure 10H]. This workflow effectively combines knowledge extraction with practical experimentation, providing a rapid, efficient, and high-precision alternative for catalyst optimization.

All in all, the roles of generative model in SAC design include generating new structures, optimizing performance, reducing costs, and enabling inverse design. GANs show immense potential in enhancing the efficiency and precision of the entire “design-screen-optimize” process in materials science. However, their poor extrapolation capability in material design limits accurate prediction of properties outside the training data scope, especially for phenomena with unclear physical mechanisms, where these models may lack in-depth explanations.