DFT/ML Method-driven Design Of The Structure Of Electrocatalyst For Carbon Dioxide Reduction Reaction

The electrochemical driven by renewable electricity reduction of carbon dioxide（CO₂）into high-value-added carbon-containing products is the key to solve energy and environmental problems and achieve the strategic goal of"emission peak and carbon neutrality".However,due to the nearly potential difference between various CO₂ reduction products,it is difficult to selectively convert CO₂ into specific products.Dual metal site catalysts（DMSCs）with tunable microstructure and complex functions have been proved to be high-efficiency catalysts to improve CO₂ reduction reaction（CO₂RR）activity and selectivity due to their particular bimetallic atom synergy effect.However,there are a large number of DMSCs,thus,the number of potential combinations to form DMSCs is bulky,which leads to the inefficiency of designing CO₂RR electrocatalysts by using traditional density functional theory（DFT）calculation or experimental methods.Therefore,it is still a huge challenge to efficiently screen the appropriate DMSCs with catalytic potential for CO₂RR.Based on these challenges,in this work,we use the combination method of DFT calculation and machine learning（ML）to construct a two round ML framework using Pearson correlation coefficient to eliminate feature variables,and construct an active learning（AL）framework using symbolic transformer to fit features.A total of 1500 DMSCs were used to predict the catalytic performance of CO₂RR to CO or HCOOH,and the relationship between bimetallic types,physicochemical properties of active sites and their electrocatalytic CO₂RR performance was explored.Four electrocatalysts with high efficiency for CO₂RR to CO and three electrocatalysts with high efficiency for CO₂RR to HCOOH were selected.The specific research contents are as follows:1.DFT/ML method-driven design of dual metal site catalysts for electrochemical CO₂ reduction to CO:the complete path of catalytic CO₂RR to CO for 98 DMSCs was calculated by DFT.The limiting potentials in the path was used as the prediction target to construct ML dataset,and the accurate prediction of catalytic performance of 1120 DMSCs in the prediction set was realized.The results shown that the optimal number of descriptors in the first round of ML is 12,and the gradient boosting regression（GBR）has the highest accuracy（RMSE=0.15V,R²=0.92）.In the second round of ML,charge transfer number and bimetallic bond length were added to further improve the performance of the GBR model（RMSE=0.08V,R²=0.98）.Two rounds of ML models complement each other,which can accurately predict the catalytic performance of DMSCs（less than 0.07 V error）while reducing the calculation of DFT.Finally,four potential electrocatalysts（Mn-Ru、Mn-Os、Zn-Ru and Co-Au-N₆-Gra-Model 3）with high activity and high selectivity for CO₂RR to CO were proposed by combining the calculation results of CO₂RR and HER.2.Active Learning Accelerating to Screen electrocatalyst for CO₂Reduction to HCOOH:an AL framework was constructed based on the DFT/ML method,and three iteration loops of AL were performed for CO₂RR to HCOOH,CO₂RR to CO and HER,achieving accurate prediction of catalytic performance of 282 DMSCs in the predicted set.The results shown that the GBR algorithm has the highest prediction accuracy and is suitable for the AL framework.By continuously increasing the DFT calculated values of 16DMSCs in each iteration loop,we reduced the uncertainty of the ML model to less than 0.20 V in the third iteration loop of AL.Among 42 DMSCs calculated by three iteration loops of AL,29 DMSCs with CO₂RR to HCOOH were selected,and the success rate was up to 70%.After the third iteration loop of AL,the prediction performance of the model was verified,and it was found that the error of ML prediction results was less than 0.20 V.Based on SHAP and PDP correlation analysis,it was found that most of the top nine descriptors of feature importance were compound descriptors,especially the covalent radius difference between bimetallic atoms,which proved that the synergistic effect between bimetallic atoms played a crucial role in activity of CO₂RR.Finally,we selected three DMSCs（Pt-Bi、Hf-Pb、As-Sb-N₆-Gra）with high activity and high HCOOH selectivity while maintaining kinetic and thermodynamic stability.In addition,we found that the ML model with good universality,which can be extended to other electrocatalysts for CO₂RR in different types and coordination environments.