Anh Khoa Augustin LU is a Senior Researcher at the Quantum Materials Simulation Group, WPI-MANA, NIMS since 2026. He previously served as an Assistant Professor in the Department of Materials Engineering at The University of Tokyo (2024–2026). His professional background includes postdoctoral research positions at NIMS (2021–2024) and at AIST (2017–2021). He received his Ph.D. from KU Leuven and imec in 2017, specializing in the ab initio modeling of novel 2D materials and interfaces. His research focuses on computational materials science, specifically atomistics simulations with density functional theory (DFT) and machine learning, using equivariant graph neural networks (SevenNet) and unsupervised learning-based structure analysis methods, to investigate complex structures and phenomena in semiconductor devices, oxides, 2D materials, and metallic glasses.

Machine learning potentials are becoming a tool of choice for atomistic simulations, offering near first-principles accuracy at a fraction of the computational cost. However, standard constant-charge models fail to capture the dynamical response of materials under an applied electric field. Previous approaches addressed this by incorporating force contribution from the electric field via a secondary neural network trained specifically for Born effective charge (BEC) tensor predictions [1]. However, running two distinct neural networks is computationally heavy and requires training separate models. System sizes have typically been limited to around 1,000 atoms per cell [2,3].

To overcome these bottlenecks, we extended SevenNet, a graph neural network based on NequIP that guarantees feature equivariance throughout its layers [4]. Using irreducible representations provides a natural framework for predicting the BEC tensor, so we modified the SevenNet architecture to create SevenNet-Polar, a multitask model capable of simultaneously predicting energy, forces, stress, and the BEC tensor [5]. We trained models on existing datasets of defect-laden ZrO₂ and Li₃PO₄ and achieved root mean squared errors (RMSE) below 1.0 meV/atom for energy, 12 meV/Å for forces, 0.05 GPa for stress, and 0.003 e for the BEC tensor, as shown in Fig. 1.

In addition to the high accuracy, we developed an interface for LAMMPS to run molecular dynamics simulations under applied electric field. A simulation speed of exceeding 1 ns/day was achieved for systems of up to 3,000 atoms on a single GPU and up to 100,000 on 64 GPUs. Simulations of systems of more than 5 million atoms under an electric field were also demonstrated. In this poster presentation, we will discuss the scaling laws governing the RMSE of each predicted quantity and show an application to ZrO₂ grain boundaries. Our methodology improves the accuracy of charge-aware machine learning force fields while making large-scale molecular dynamics under electric fields possible even with limited computational resources.

1. K. Shimizu, R. Otsuka, M. Hara, E. Minamitani, S. Watanabe, STAM Methods 3, 2289659 (2023), DOI: 10.1080/27660400.2023. 2253135
2. S. Minami, A. Kutana, R. Jinnouchi, R. Asahi, Phys. Rev. Materials 9, 103802 (2025), DOI: 10.1103/PhysRevMaterials.9.103802
3. A. Lu, N. Maekawa, A. Ikeda, H. Masuda, H. Yoshida, S. Watanabe, Phys. Rev. Materials 10, 066601 (2026), DOI: 10.1103/jcsd-dbl2
4. Y. Park, J. Kim, S. Hwang, S. Han, J. Chem. Theory Comput. 20, 4857 (2024), DOI: 10.1021/acs.jctc.4c00190
5. A. Lu, S. Arai, Y. Park, S. Han, T. Miyazaki, S. Watanabe, arXiv:2607.14827 (2026), DOI: 10.48550/arXiv.2607.14827