Machine Learning for Hydrogen Storage in Boron‐Based Materials: Key Descriptors and Data Constraints

ABSTRACT

The transition to a sustainable energy economy critically depends on the development of safe, efficient, and cost‐effective hydrogen storage solutions. Boron‐based materials are promising solid‐state hydrogen storage candidates with high gravimetric capacity, but their use remains limited by kinetic and thermodynamic challenges. This study employs a data‐driven approach with machine learning (ML) to systematically analyze the structure–property–performance relationships in these systems. A curated dataset comprising 1073 experimental entries from 20 articles (2000–2023) was analyzed using regularized linear regression (Ridge, Lasso, and Elastic Net), random forest (RF), and gradient boosting (GB) algorithms, using an article‐wise cross‐validation scheme that kept all data from the same publication within a single fold, preventing information leakage. The linear models were unable to describe the adsorption behavior adequately, and expanding the descriptor set did not improve them, indicating an intrinsic limitation of the linear formulation. The nonlinear ensembles performed clearly better, indicating that the underlying relationships are nonlinear, though their overall accuracy was still modest. Despite its greater complexity, GB did not outperform RF, which is likely related to the outlier‐prone and uneven distribution of the data. SHAP analysis showed that both ensembles identified the same leading descriptors, namely the preparation method and adsorption pressure, followed by temperature, surface area, and boron and carbon content. The direction of their effects agreed with known adsorption behavior. In addition, the incomplete reporting of key structural descriptors such as surface area and pore volume limited the predictive capacity of all models. Still, the results provide a reproducible framework for predictive modeling of hydrogen storage and identify the material features that most strongly affect storage capacity.