MACHINE LEARNING-GUIDED DESIGN AND OPTIMIZATION OF HYDROGEN EVOLUTION REACTION (HER) CATALYSTS FOR EFFICIENT HYDROGEN PRODUCTION

Abstract

The development of efficient and cost-effective electrocatalysts for the hydrogen evolution reaction (HER) is critical for enabling sustainable green hydrogen production through water electrolysis. However, traditional catalyst discovery relying on density functional theory (DFT) calculations and experimental trial-and-error is           computationally intensive and time-consuming. Herein, we present a machine learning-guided framework for the rapid design and optimization of HER catalysts within the MXene and MBene families, two-dimensional materials that have emerged as promising alternatives to noble metals. A comprehensive dataset comprising 285 materials was curated from computational databases, with 28 descriptors spanning compositional, electronic, structural, and thermodynamic categories. Eight supervised regression algorithms were systematically evaluated, with XGBoost emerging as the optimal model after hyperparameter optimization, achieving exceptional predictive performance (R² = 0.90, MAE = 0.06 eV, RMSE = 0.12 eV) on test data. SHAP (SHapley Additive exPlanations) analysis revealed that electronic descriptors, particularly metal valence electrons and electronegativity difference between metal and non-metal components, dominate predictions, accounting for approximately 60-70% of the model's predictive power. Surprisingly, traditionally emphasized descriptors such as d-band center showed minimal importance, suggesting that for MXenes and MBenes, valence electron configuration subsumes the predictive capability of more complex electronic features. Feature reduction from 28 to 20 optimal descriptors improved both model performance and interpretability. High-throughput screening identified several promising molybdenum-based candidates with near-optimal hydrogen adsorption free energies (ΔG_H* approaching 0 eV), including Mo₅B₂ (-0.54 eV), MoC₂ (-0.33 eV), Mo₂C (-0.36 eV), and Mo₂N (-0.34 eV), all earth-abundant materials warranting urgent experimental validation as potential platinum alternatives. This work establishes a robust, interpretable machine learning framework that achieves 90% predictive accuracy using only readily available elemental descriptors, bypassing the need for computationally intensive DFT calculations in initial screening phases. The methodology provides a powerful tool for accelerating the discovery of next-generation HER electrocatalysts, directly supporting the global transition toward sustainable hydrogen production and net-zero carbon emissions.