Knowledge distillation of neural network potential for molecular crystals

Abstract

Organic molecular crystals exhibit various functions due to their diverse molecular structures and arrangements. Computational approaches are necessary to explore novel molecular crystals from the material space, but quantum chemical calculations are costly and time-consuming. Neural network potentials (NNPs), trained on vast amounts of data, have recently gained attention for their ability to perform energy calculations with accuracy comparable to quantum chemical methods at high speed. However, NNPs trained on datasets primarily consisting of inorganic crystals, such as the Materials Project, may introduce bias when applied to organic molecular crystals. This study investigates the strategies to improve the accuracy of a pre-trained NNP for organic molecular crystals by distilling knowledge from a teacher model. The most effective knowledge transfer was achieved when fine-tuning using only soft targets, i.e., the teacher model's inference values. As the ratio of hard target loss increased, the efficiency of knowledge transfer decreased, leading to overfitting. As a proof of concept, the NNP created through knowledge distillation was used to predict elastic properties, resulting in improved accuracy compared to the pre-trained model.