Toward Reliable and Transferable Machine Learning Potentials: Uniform Training by Overcoming Sampling Bias

The neural network interatomic potential (NNP) is anticipated to be a promising next-generation atomic potential for its self-learning capability and universal mathematical structure. While various examples demonstrate the usefulness of NNPs, we find that the NNP suffers from highly inhomogeneous feature-space sampling in the training set. As a result, underrepresented atomic configurations, often critical for simulations, cause large errors even though they are included in the training set. Using the Gaussian density function (GDF) that quantifies the sparsity of training points, we propose a weighting scheme that can effectively rectify the sampling bias. Various examples confirm that GDF weighting significantly improves the reliability and transferability of NNPs compared to the conventional training method, which is attributed to accurate mapping of atomic energies. By addressing a detrimental problem that is inherent in every machine learning potential, the present work will extend the application range of the machine learning potential.