Fujitsu achieves high-precision, long-duration molecular dynamics simulation for all-solid-state battery interphases with over 100,000 atoms

Fujitsu developed a knowledge distillation technique (Figure 1) to precisely train NNPs with a faster multi-layer perceptron (MLP) [7] architecture. This is achieved by transferring knowledge from computationally slower, but knowledge-rich, GNN-based published NNPs. This approach allows MLP-based NNPs to leverage extensive published NNP knowledge and specialized material structure insights, enabling stable, high-speed, long-duration MD simulations for large-scale systems exceeding 100,000 atoms.

When applied to a next-generation all-solid-state battery interface (127,296 atoms), Fujitsu confirmed stable, 10-nanosecond MD simulations in approximately one week (Figure 2(b)). This enabled structural analysis of the SEI, critical for battery performance and previously unachievable with existing MD simulations. SEI defines the charge-discharge cycle life and safety of all-solid-state batteries, and understanding its atomic-level formation and stability is crucial. This technology is expected to accelerate the development of a method to control SEI formation by elucidating its previously unknown atomic-level processes.

NNP-based MD simulations have recently gained traction for rapidly and accurately simulating material properties at the atomic level. Published NNPs, trained on millions of diverse material data points, are increasingly utilized.
However, material structure collapse during simulation has been an ongoing key challenge with published NNP-based MD simulations, especially for complex materials like all-solid-state batteries. Furthermore, many published NNPs, trained on extensive datasets, employ graph neural networks (GNNs) [6]. While expressive, GNNs are computationally slow, taking over a year for long-duration simulations of large-scale systems exceeding 100,000 atoms, rendering them impractical. This new technology attempts to address these challenges.