Interpretable machine learning enables high performance of magnesium borohydride hydrogen storage system

Abstract

Aiming to provide optimal solutions to the sluggish kinetics of Mg(BH₄)₂, this study proposes, for the first time, a novel machine learning model to predict dehydrogenation behaviors of modified Mg(BH₄)₂. Notably, numerous data points are collected from temperature-programmed, isothermal, and cyclic dehydrogenation behaviors, a neural network model is proposed by using multi-head attention mechanisms, which exhibits the highest predictive performance compared to traditional machine learning models. The study also ranks different variables influencing dehydrogenation processes, employing interpretable analysis to identify critical variable thresholds, offering guidance for the experimental parameter design. The model can also be adapted to scenarios involving co-doping of hydrides and catalysts in Mg(BH₄)₂ system and proved high accuracy and scalability in predicting dehydrogenation curves under diverse conditions. Employing the model, performance predictions for a series of undeveloped Mg(BH₄)₂ co-doping systems can be made, and superior dehydrogenation catalytic effects of fluorinated graphite (FGi) are uncovered. Real-world experimental validation of the optimal Mg(BH₄)₂-LiBH₄-FGi system confirms consistency with model predictions, and performance enhancement attributes to experimental parameter optimization. Further characterizations provide mechanistic insights into the synergistic interactions of FGi and LiBH₄. This work paves the way for advancing utilization of machine learning in the high-capacity hydrogen storage field.