Machine Learning Assisted Optimization Design Of MXenes Hydrogen Evolution Reaction Catalysts

Two-dimensional MXenes have emerged as a promising class of materials for catalyzing the hydrogen evolution reaction(HER),as they offer numerous advantages such as excellent surface hydrophilicity,a large specific surface area,and excellent electrical conductivity.Doping single transition metal atoms(STM)onto the surface of MXenes is a promising strategy to regulate the hydrogen adsorption Gibbs free energy(ΔGH)and further expand the pool of potential catalysts.However,traditional high-throughput calculations for identifying suitable doping catalysts can be extremely time-consuming.Therefore,there is a pressing need to develop efficient methods for predicting and screening MXenes-STM materials with low hydrogen adsorption Gibbs free energy to facilitate accurate research and development.Unfortunately,current research on doping MXenesSTM is still in its initial stage,and there is no effective descriptor or construction method for predicting Gibbs free energy of hydrogen adsorption,nor a general prediction model for efficient screening.This lack of knowledge severely hinders the exploration of new MXenes materials.In this study,we aimed to develop machine learning software suitable for materials science applications,explore its HER descriptor combined with first-principles calculation,reveal the regulatory mechanism of MXenes catalytic activity under the influence of electronic structure and size factors,establish a unified machine learning model,and explore potential catalysts for HER.(1)To address the construction requirements for new descriptors,we developed a general knowledge embedded symbolic regression algorithm(SRDSC).This algorithm incorporates dimensional computation and descriptor binding mechanisms into the symbolic regression modeling process to generate concise and high-precision physical expressions.To verify its effectiveness,we used a binary compound band gap data set as an example and employed symbolic regression combined with machine learning descriptor screening to construct new descriptors.Our approach successfully predicted the band gap(Eg)of NaCl-type and cubic ZnStype compounds.The developed SR-DSC algorithm and related feature preprocessing tools enable descriptor screening and cover the entire modeling process,providing valuable tool support for future studies.(2)The hydrogen evolution mechanism of electron local MXenes doping systems(Ti2CO2-STM,Zr2CO2-STM,Ta2CO2-STM)was studied.We investigated the influence of electron and size factors on catalytic activity and established a new descriptor using symbolic regression,specifically φ=Ef/dM1-O-rM1,which predicts the Gibbs free energy of hydrogen adsorption in MXenes-STM systems with electron localization.To ensure the hydrogen adsorption Gibbs free energy and system stability,we selected Ti2CO2-W with | ΔGH |<0.1 eV as a suitable MXenes material.In this study,we developed a two-channel screening optimization process for efficient high-throughput computation and descriptor prediction.(3)To address the imperfect hydrogen evolution mechanism in electron nonlocal MXenes-STM systems,we discussed the regulation rule of STM doping on catalytic activity for the hydrogen evolution reaction.Our study focuses on the twodimensional Mo2CO2 system,which exhibits non-local electrons that lead to next neighbor bonding and affect the adsorption of Gibbs free energy of hydrogen.We employed machine learning to screen descriptors containing Fermi energy level and local structure information and identified five descriptors for catalytic activity of hydrogen evolution that reveal the regulatory mechanism of electron doping in nonlocal systems.These descriptors were successfully applied to predict the behavior of W2CO2-STM in hydrogen evolution.(4)Efficient screening and prediction of Gibbs free energy for hydrogen adsorption by MXenes was achieved through the establishment of a CGGRU crystal graph cyclic network structure and a prediction model targeting MXenes unit atomic energy.Using this model framework and the | ΔGH |<0.1 eV standard,we were able to select 1906 potential catalysts from 75,816 MXenes composition proportions,achieving a screening efficiency 40 times higher than exhaustive calculation.This approach provides an efficient method for the screening of MXenes materials.Our study focused on developing a machine learning tool set that is suitable for material researchers,with a specific focus on algorithms that can reduce the threshold for knowledge extraction and new material prediction.In the context of MXenes material application,we developed a dual screening method using descriptors and neural network models to explore potential excellent MXenes-based hydrogen evolution catalysts.By leveraging these tools,we were able to streamline the screening process and improve our ability to predict the properties of new materials.