Machine learning based prediction of metal hydrides for hydrogen storage, part II: Prediction of material class

Abstract The openly available dataset on hydrogen storage materials provided by the US Department of Energy was used to predict the optimal materials class of metal hydrides based on the desired properties, which included hydrogen-weight percent, heat of formation and operating temperature and pressure. We performed correlation and statistical analyses to investigate the statistical characteristics of each numeric features. We employed four classification algorithms: multiclass logistic regression, multiclass decision forest, multiclass decision jungle and multiclass neural network. Feature importance analysis was carried out to investigate how each classifier utilises the information available in the dataset. In overall, multiclass neural network classifier had better classification performance obtaining an accuracy of 0.80. The results suggest that the complex material class, followed by Mg is applicable for the most wide range of operating temperatures. Positive correlation was found between hydrogen weight percent, heat of formation and temperature, suggesting that the maximum hydrogen weight percent would be achieved in the complex material class operated at a high temperature.