On Transferability of Machine Learning Force Fields: A Case Study on Silicon

In this article, we present a systematic study in developing machine learning interatomic potential (MLIAP) for elemental silicon. Previously, many of MLIAP applications involve fitting models to small or localized training sets, which do not allow for transferability. The need for transferability of MLIAP is crucial in sampling the vast phase space in the application of crystal structure prediction. To address for this need, we indicate that two primary factors can have major influences on the transferability of the MLIAP. First, transferability depends strongly on the diversity of the training data set. In particular, the use of a random structures generating algorithm based on crystal symmetry is a suitable methodology for generating a training data set with sufficient diversity. Second, we introduce a framework that allows Gaussian symmetry functions and bispectrum coefficients descriptors to be fitted with generalized linear regression or neural network. According to our substantial benchmark, the best accuracy trained on the diverse data set can be achieved by using a flexible machine learning method coupled with highly orthogonal descriptors. Therefore, we will show that neural network potential fitting with bispectrum coefficients as the descriptor is a feasible method for obtaining high-quality and transferable MLIAP.