XPEAK: An XRD-Driven Machine Learning Platform for Predicting the Catalyst-Enhanced Dehydrogenation Peak Temperature of MgH₂

Abstract

MgH₂ offers high hydrogen capacity but suffers from an excessively high desorption temperature, which severely restricts its practical application. Developing effective catalysts is essential, yet their structural complexity makes quantitative description and performance prediction notoriously difficult. To overcome the long-standing challenge of catalyst representation, we extracted structural fingerprints directly from standardized X-ray Diffraction (XRD) patterns using dimensionality reduction. These XRD features integrated with catalyst mass fraction and heating rate, were employed to build the machine learning models in the field of MgH₂ catalysis. To predict the dehydrogenation peak temperature (Tp) and its variation (ΔTp), we compiled a curated dataset from 420 experimental publications, containing over 2,000 records. Ensemble regression models were trained separately on experimental and ICDD-based XRD data. Model interpretation using SHAP highlights that XRD-derived features dominate prediction accuracy, while catalyst loading and heating rate show nonlinear but interpretable effects. Based on ICDD-based XRD data, we developed a high-throughput screening workflow coupled with t-Distributed Stochastic Neighbor Embedding visualization, enabling the rapid evaluation of over 100,000 ICDD XRD entries and the identification of 12 promising catalyst candidates (Tp < 230 °C). Based on experimental XRD data, we established a predictive framework for dehydrogenation peak temperatures, capable of integrating user-provided catalyst XRD data. Importantly, we implemented this workflow in XPEAK (an XRD-driven machine learning platform for predicting peak temperatures, https://cat-metalhy.cpolar.top/), the first platform providing the research community with practical AI-assisted tool for catalyst screening, performance prediction, and design for MgH₂ dehydrogenation. This work demonstrates how XRD fingerprints can resolve the bottleneck of catalyst description and establish a new pathway for data-driven catalyst design.