Machine-learning descriptor search on the density of states profile of bimetallic alloy systems and comparison with the d-band center theory.

Abstract

In this study, the electronic density of states (DOSs) calculated with density functional theory (DFT) were analyzed by the machine-learning techniques. More than 400 pure metal and bimetallic alloy systems were calculated with DFT, and obtained the surface DOSs and the CH3 adsorption energy (Ead). By fitting the Gaussian functions to the DOS, multiple descriptors, such as the Gaussian peak positions, heights, and widths were extracted. Several regression methods, such as the least absolute shrinkage of selection operator (LASSO), random-forest, gradient-boosting, and extra-tree were used to find the relationship between these descriptors and the Ead. The results show that the energy position of the peaks in the d-projected DOS is the most important descriptor, in agreement with the previously known d-band center theory. It was also shown that the peak position in d-projected DOS improves the regression model in addition to the d-band center, since it reduces the regression error.