Transfer Learning of Potential Energy Surfaces for Efficient Atomistic Modeling of Doping and Alloy

This letter proposes a transfer learning (TL) method to generate neural network (NN) database to model doping and alloy. By leveraging the valuable potential energy surface (PES) information already available in source system and similarities between source and target systems, the proposed TL successfully reduces computational cost by several orders of magnitude, while keeping ab-initio level high accuracy. We show that it is generally applicable to model ${p}$ -type, ${n}$ -type, and alloy atomic substitutions.