Machine learning enabled customization of performance-oriented hydrogen storage materials for fuel cell systems

Hydrogen storage materials with different crystal configurations have been extensively investigated for hydrogen promotion. To escape the dilemma of traditional trial-and-error composition optimization, in this work, efficient implicit/explicit features-based machine learning (ML) was applied for the first time to typical metal hydrides with the newly-constructed proprietary dataset. Excitingly, the most pivotal capacity-affecting factors (MeanIonicChar value/Fe content) were identified through feature importance ranking, facilitating efficient capacity estimation and formulation of high-capacity compositions. Subsequently, ML-based proactive properties scanning and composition customization were performed for fuel cell hydrogen feeding system. Generally, the measured hydrogen storage properties exhibit satisfactory accuracy and a validation relationship with the ML-based predicted values. In addition, an intrinsic link between atomic occupancy randomness and pressure-composition-temperature slope was revealed by theoretical calculations. Among the alloys developed from the advanced paradigms, Ti0.9Zr0.12Mn1.2Cr0.55(VFe)0.25 offers all-round properties (1.90 wt% / 127.30 kg H2/m3 in saturation) and overwhelming cost-effectiveness compared with the reported alloys at the moderate temperature and pressure level. In summary, ML-based composition customization pathways avoid substantial experimental investments and provide a novel option for efficient acquisition of high-performance hydrogen storage materials.