Data-Driven Catalyst Design: A Machine Learning Approach to Predicting Electrocatalytic Performance in Hydrogen Evolution and Oxygen Evolution Reactions (2412.12846v1)

Abstract: The transition to sustainable green hydrogen production demands innovative electrocatalyst design strategies that can overcome current technological limitations. This study introduces a comprehensive data-driven approach to predicting and understanding catalytic performance for Hydrogen Evolution Reaction (HER) and Oxygen Evolution Reaction (OER) using advanced machine learning methodologies. By usimg a dataset of 16,226 data points from the Catalysis-hub database, we developed a novel stacking ensemble model that integrates Random Forest, XGBoost, and Support Vector Regression to predict Gibbs free energy of adsorption across diverse bimetallic alloy surfaces. Our innovative feature engineering strategy combined Matminer-based compositional analysis, Principal Component Analysis for adsorption site related features, and correlation screening to generate robust predictive descriptors. The machine learning model demonstrated exceptional predictive capabilities, achieving R^2 values of 0.98 for HER and 0.94 for OER, with Mean Absolute Error values of 0.251 and 0.121, respectively. Shapley Additive Explanations (SHAP) analysis revealed critical insights into the complex interplay of compositional, structural, and electronic features governing catalytic performance. The research provides a powerful computational framework for accelerating electrocatalyst design, offering unprecedented insights into the fundamental properties that drive hydrogen evolution and oxygen evolution reactions. By bridging advanced machine learning techniques with fundamental electrochemical principles, this study presents a transformative approach to developing cost-effective, high-performance catalysts for sustainable hydrogen production.