Dataset Expansion Strategy for Interatomic Potential by Descriptor-informed Active Learning

Abstract

In Automotive industry, reducing carbon dioxide emissions and achieving circular economy such as materials recycle are strongly necessary. The automotive materials become more diverse and complex by vehicle electrification and renewable raw materials usage. Materials Informatics (MI) and atomic scale analysis are expected to elicit high performances of many materials used in a car. Fast development of precise interatomic potentials which calculate the material properties is important for applying atomic scale analysis and MI to develop automotive materials. For efficient collection of training data for developing the interatomic potential, we investigate a method to generate new training data from the descriptors and a strategy to suggest new descriptor for the training data. We reveal that Genetic Algorithm & Gradient Descent Hybrid method is effective for generating new training data by optimizing with descriptors of existing training data. Moreover, we achieved expand dataset by sampling with low density area & k-nearest neighbor strategy.