Work

Predicting hydrogen storage capacity of metal hydrides using novel imputation techniques and tree-based machine learning models

Year: 2025

Type: article

Abstract: Hydrogen is emerging as a crucial clean energy carrier, with the potential to decarbonize transportation, industry, and power generation. Among various storage methods, metal hydrides are particularly promising due to their high volumetric density, safety, and reversibility under moderate conditions. However, experimental research to optimize hydrogen absorption in metal hydrides is time-consuming and costly, motivating the use of alternative approaches to accelerate material development. In this study, we apply machine learning (ML) techniques to predict hydrogen storage capacities in metal hydrides based on compositional and structural features. The analysis used an experimental dataset covering eight material classes. To address the limited sample size, imputation techniques were applied to generate additional data points, enhancing model training without discarding existing measurements. Feature engineering and average importance analysis were used to identify the most influential variables, creating a structured input for ML models. The trained model achieved a preliminary coefficient of determination ( R 2 ) of 0.92 and a mean absolute error (MAE) of 0.043. Postprocessing further improved performance, raising R 2 to 0.99 and reducing MAE to 0.0052. While the small dataset limits generalizability, the combination of effective imputation, feature selection, and ML modeling achieves high predictive accuracy, as validated using the LIME and SHAP frameworks. These results highlight the potential of integrating ML techniques into materials science, providing a promising strategy to accelerate the development of next-generation hydrogen storage solutions while motivating further experimental validation. • Predicting hydrogen storage in metal hydrides using ML (R² = 0.99, MAE = 0.0052). • Addressing limited datasets through imputation and feature engineering. • Identifying influential compositional and structural features via ML. • Validating model interpretability using LIME and SHAP frameworks. • Accelerating metal hydride development beyond time-consuming experiments.

Source: Computational Materials Science

DOI: 10.1016/j.commatsci.2025.114390

Authors: Zaid Allal, Hassan N. Noura, Flavien Vernier, Ola Salman, Khaled Chahine

Institutions: Université Savoie Mont Blanc, American University of Beirut, American University of the Middle East

Language: English

Topic: Machine Learning in Materials Science

Subfield: Materials Chemistry

Field: Materials Science

Domain: Physical Sciences

SDGs: Affordable and clean energy, Industry, innovation and infrastructure

Open Access status: hybrid

Estimated APC paid: $3,290