Explainable machine learning and physics-constrained optimization of chalcogen catalysts for sustainable hydrogen production
TL;DR
An integrated machine learning and optimization framework, physics-constrained optimization (PCO) and trust-region Bayesian optimization (TRBO), for engineering of chalcogen-based electrocatalysts, offering a physically consistent, reliable, and scalable route for the discovery of new catalysts and sustainable hydrogen production.
Abstract
Summary The accelerating demand for clean energy has intensified research on sustainable hydrogen production, where efficient chalcogen catalyst design remains a major challenge. Most of the existing studies are, however, only concerned with predictive modeling, neglecting physical feasibility and consistency in the optimization process. To overcome this, this study proposes an integrated machine learning and optimization framework, physics-constrained optimization (PCO) and trust-region Bayesian optimization (TRBO), for engineering of chalcogen-based electrocatalysts. The proposed framework takes into account domain-specific physical band gap constraints, such as physically stable formation energies and realistic density bounds, to guarantee physically meaningful and practically feasible solutions. The XGBoost model trained with engineered Magpie descriptors has demonstrated excellent predictive accuracy with R2 of 0.961 along with low prediction error. PCO also showed better results with the highest band gap of 2.406 eV as compared with the 2.276 eV of TRBO. SHapley Additive exPlanations (SHAP) analysis also revealed the most influential descriptors that govern the behavior of the band gap, enhancing the interpretability of the model. In summary, the proposed framework offers a physically consistent, reliable, and scalable route for the discovery of new catalysts and sustainable hydrogen production.