Accelerating the design and optimization of catalysts for the hydrogen evolution reaction in transition metal phosphides using machine learning.

• The predictive ability of ΔG H for TMPs is enhanced by the integrated algorithm. • In terms of data sources, simply combining experimental or calculated data. • Feature importance analysis includes global and local feature importance. Transition metal phosphides (TMPs) are potential substitutes for platinum group metal catalysts in hydrogen evolution reactions (HER) for sustainable energy. In this work, a novel machine learning (ML)-based technique for HER performance prediction of TMPs is presented. ΔG H , which assesses catalytic activity, was utilized as the expected target value and data from sources like collected papers and self-calculated data were also employed. The random forest (RFR) technique produced an R2 of 0.89 on the test set, while non-tree ML algorithms were utilized for learning and feedback. Additionally, TMPs of ΔG H were predicted using an original Ensemble learning technique. The ensemble learning method significantly increased the model's predictive effect and strengthened the model's stability and resilience, resulting in an R2 >0.85 on the test set for the tree algorithm model. The SHAP (SHapley Additive exPlanations) algorithm and RFR showed that the geometric means of atomic volume, first ionization energy, and p-orbital electron number significantly impact free energy performance evaluation by tree ML algorithms. These findings can effectively guide the design and development of new products. In conclusion, this approach offers a useful tool to assess TMPs' performance in HER studies. [Display omitted] [ABSTRACT FROM AUTHOR]