[S-CG49] Innovation through the Integration of Solid Earth Science and Materials Science

[SCG49-01] Machine-learning interatomic potentials in materials science

★Invited Papers

*Kobayashi Ryo¹ (1.Nagoya Institute of Technology)

Keywords:Machine-learning potential, Molecular dynamics, Defects

In materials science, the identification of atomic and crystal structures and the accurate prediction of electronic, optical, magnetic, thermal, and mechanical properties of a given structure are important issues. In order to accurately predict the mechanical properties of materials, it is necessary to accurately describe the structure and dynamic behavior of defects in materials. The difficulty of analyzing the details of the dynamic behavior of atoms and molecules using experimental methods alone has led to the development of molecular dynamics (MD) simulations of atomic- and molecular-scale dynamics and many interatomic potentials that reproduce well the characters of atoms and behaviors of atomic bonds. However, it has been difficult for these classical potentials to accurately reproduce chemical reactions involving quantum mechanical state changes in many cases. Under such a circumstance, many machine-learning type potentials have been developed with the recent development of AI and machine learning. In the early machine-learning type potentials (Behler-Parrinello type), the coordinates of atoms and molecules were transformed into invariant descriptors for translation and rotation, called symmetry functions, and assigned to a neural network (NN), which successfully predicted the translation- and rotation-invariant energies. Subsequently, many descriptors and machine learning models were proposed, and highly accurate potentials for materials consisting of relatively few elements were successfully generated. Recently, machine learning potentials with higher accuracy, generalization performance, and efficiency have been developed by using graphs to represent the environment around atoms, keeping quantities that are equivariant to translation and rotation as internal variables, and using a structure in which the model size does not change with changes in the number of elements. Using a large amount of first-principles data, general-purpose machine learning potentials have also been developed that can be used for almost any combination of elements. Our group has developed active learning schemes for learning machine learning potentials for complex atomic structures, such as dislocation cores, using a equivariant-type graph NN potential, as well as methods for constructing more accurate potentials by introducing an attention mechanism used in large-scale language models (LLMs) to the graph NN potentials.
The talk will discuss the development of machine learning potentials to date and some examples, as well as potential improvements and applications in the earth sciences.