Machine learning prediction of hydrogen adsorption energy on platinum nanoclusters: A comparative study of SOAP descriptors

Abstract

Hydrogen binding energy in metal materials is of high significance in the hydrogen storage as well as the hydrogen evolution reaction of electrocatalysis. In this work, the datasets (more than 9000 data) of hydrogen adsorbed on Pt nanoclusters with different sizes are obtained by first-principles calculations. Data analysis shows that the binding strength of hydrogen with Pt is closely relevant to the local structures of the adsorption sites. The local features of the distance between the platinum and hydrogen and the size of the nanoclusters are supplemented to the Smooth Overlap of Atomic Positions descriptors to fit and predict the adsorption energies of hydrogen on different Pt nano-structures by performing the machine learning method. Gaussian Process Regression (GPR) and Random Forest Regressor (RFR) are used to construct the prediction model of hydrogen binding energies and it is found the R2 of test set is improved from 0.63 to 0.78 with modified descriptors. By applying it into other nanoclusters, the MAE of the prediction model is 0.08 eV, which exhibits high accuracy of the hydrogen adsorption energy. Our model can be easily extended to the prediction of hydrogen adsorption energy of other materials with affordable computational cost and accuracy, which would be helpful for the structural design of high-performance catalysts.