Prediction of Pressure-Composition-Temperature Curves of AB(2)-Type Hydrogen Storage Alloys by Machine Learning

Abstract

Pressure-composition-temperature (PCT) curves for hydrogen absorption and desorption of AB(2)-type hydrogen storage alloys at arbitrary temperatures are predicted by three machine learning models such as random forest, K-nearest neighbor and deep neural network (DNN). Two data generation methods are adopted to increase the number of data points. A new form of the PCT curve functions is suggested to fit experimental data, which greatly helps improve the prediction accuracy. A van't Hoff type equation is used to generate unmeasured temperature data, which improves the model performance on the PCT behavior at various temperatures. The results indicate that a DNN is the best model for predicting the PCT behavior with a high average correlation value R-2 = 0.93070.