Machine Learning Accelerates The Controlled Construction And Catalytic Performance Of MXenes Based Heterojunctions

The effective use of integrated solar-hydrogen cogeneration systems contributes to the sustainable and efficient development of clean energy.MXenes-based heterostructures with large specific surface area and ultra-high conductivity have good catalytic properties in oxygen reduction reactions（ORR）and oxygen reduction reactions（OER）.Combining machine learning with first principles methods can accelerate MXenes heterojunction design and improve the efficiency of energy conversion devices.Although MXenes are partially synthesized experimentally,there is a relative lack of theoretical studies and ML for catalytic descriptor construction.To address the problems of high cost and low efficiency of DFT calculations due to multiple combinations of MXenes heterojunction species,this paper focuses on the combination of density function theory and machine learning（DFT-ML）to construct heterostructures of transition metal-doped nitrogen-liganded graphene with two-dimensional MXenes,analyze their structural stability,establish catalytic descriptors,and explore their ORR and OER catalytic reaction mechanisms from the electronic structure level.Details are as follows.（1）MXenes heterojunction construction and feature parameters of machine learning.Eight heterostructures（M-N₄-Gr/V₂C）of 3d transition metal-doped nitrogen-coordinated graphene with MXenes（V₂C）with good electronic properties were established by DFT calculations.The stability of the structures was obtained by calculating the formation energy.The calculation of ORR and OER thermodynamic processes clarifies the intrinsic connection between electronic structure and catalytic activity,and the relationship between Gibbs free energy and d-band center of oxygen-containing intermediates,respectively,and overpotential is obtained to provide feature parameters for machine learning.（2）Machine learning accelerated the prediction of MXenes-based heterojunctions and catalytic performance.The screening of 3d,4d,and 5d transition metal-doped M-N₄-Gr/MXenes（Ti₂C,Nb₂C,V₂C）heterojunctions was accelerated by the DFT-ML method;among the five machine learning algorithms,the random forest algorithm（RFR）was verified to be the most ideal algorithmic model;the descriptors of transition metal intrinsic features and overpotentials（φ）were established by the feature importance share and Pearson correlation,and the ORR and OER catalytic activities were predicted using the descriptors.（3）DFT performs reverse validation of machine learning prediction results and catalytic mechanism analysis.After the comparison between the DFT calculation and the machine learning prediction data,the predicted and validated values were obtained with small errors（0.02～0.17 V and 0.03～0.10 V for ORR and OER overpotential errors,respectively）,which proved the accuracy of the machine learning prediction;the catalytic reaction mechanism was clarified by analyzing the thermodynamic processes,d-band centers,and the chemical bonding strength between the active site metal and oxygen-containing intermediates.The best ORR catalyst Ni-N₄-Gr/Nb₂C（0.31V）,the best OER catalyst Co-N₄-Gr/V₂C（0.30V）and the bifunctional ORR/OER catalyst Co-N₄-Gr/V₂C（0.45V/0.30V）were successfully screened from 78 M-N₄-Gr/MXenes catalysts for hydrogen for fuel cells and electrolytic water hydrogen production research to provide a theoretical basis.