Development of first-principles pseudopotentials and numerical atomic orbitals using machine learning and its accuracy assessment

Kawai, Hiroyuki*; Sekikawa, Takuya; Nakamura, Yoshiyuki*; Ozaki, Taisuke*; Ono, Yoshiaki*

OpenMX is a density-functional-theory based software for first-principles electronic structure calculations, and is mainly used to obtain the most stable structures and electronic states of materials. In this study, we attempted to replace the numerical data that had been created by adjusting parameters based on the creator's experience with numerical data optimized by machine learning. In particular, for Li, we succeeded in reducing the calculation error to about one-third that of the exact solution by optimizing the data with machine learning. This research provides new knowledge for improving the calculation accuracy of OpenMX and for creating numerical data for actinides, which OpenMX currently cannot handle.