Discovering elemental effects on hydrogen storage in TiMn2-type alloys using explainable machine learning

Abstract

Efficient hydrogen storage is critical for advancing a hydrogen-based energy economy, yet optimizing the interdependent properties of TiMn2-type alloys remains challenging. Here, we present a machine learning (ML) framework that integrates multi-objective optimization, experimental validation, and explainable AI (XAI) to guide alloy discovery. Trained on 508 data points across 207 compositions, our graph-based models predict key performance metrics, including equilibrium pressures, hysteresis, plateau slope, and capacity, with high accuracy. Experimental validation of five ML-designed alloys confirms the framework's reliability, with one candidate approaching the Pareto front under practical constraints. XAI analysis reveals how elemental properties influence performance, enabling interpretable predictions and identifying underexplored elements such as Nb. This work establishes a generalizable and interpretable approach for accelerating hydrogen storage alloy design.