Prediction of Atomization Energies of Au13+ Clusters through the Machine Learning Approach

Abstract

We examine a new method for predicting the atomization energies of Au13+ clusters by a nonlinear regression model using interatomic and centroid distances as descriptors to improve the efficiency of density-functional theory calculations. Learning data were created using the time-series data of atomic coordinates and Kohn-Sham energy generated by molecular-dynamics simulations. This approach predicted the atomization energies of fifteen known stable/metastable structures of Au13+ clusters well. Moreover, we found that the fitting to the test data could be markedly improved by eliminating the descriptors representing the short interatomic distance.