122nd General Meeting of the KCS

Type Symposium
Area Data-Enabled Computational Chemistry
Room No. Room 324A
Time FRI 09:25-:
Code PHYS2-2
Subject Classical neural network potential learning DFT results: a new frontier in physical chemistry
Authors Wonseok Jeong, Seungwu Han*
Materials Science and Engineering, Seoul National University, Korea
Abstract By revealing atomic trajectories, classical molecular dynamics (MD) simulations have advanced fundamental understanding on various physical and chemical processes at the atomistic level. In classical MD, the chemical bonds are approximated by interatomic potentials that are parameterized by fitting key properties to reference data. The functional form of interatomic potentials reflects the underlying bonding nature. In many materials, however, the bonding nature is rather mixed, which makes it difficult to choose a proper function type. Recently, the machine-learning (ML) potential is gaining traction as a data-driven approach to generating interatomic potentials. In contrast to traditional interatomic potentials with preset analytic functions, the ML potentials assume flexible mathematical structures such as neural network and their parameters are optimized through machine learning on extensive reference data. While NNP is getting popular, the weakness and strength of NNP are not fully understood at this moment, mainly because of its ‘black-box’ nature. Here we show that NNP suffers from inhomogeneous feature-space sampling in the training set. As a result, underrepresented atomic configurations 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 rectify the sampling bias. Various examples confirm that the GDF weighting significantly improves reliability and transferability of NNP compared to the conventional method, which is attributed to accurate mapping of atomic energies. By addressing a detrimental problem that is inherent in every ML potential, the present work will extend the application range of the ML potential.
E-mail hansw@snu.ac.kr