Machine learning potentials have emerged as a cost-effective alternative to first-principles calculations with high accuracy and high simulation speed. However, atomic charges are typically neglected, thus limiting their field of applications when an electric field is applied, which is essential in semiconductor devices. Several methods have been proposed to describe dynamical charges, but they typically have to compromise between simulation speed and accuracy. By developing an optimized architecture for message-passing graph neural network with SevenNet-Polar, we train highly accurate force field while ensuring efficient scalability, enabling molecular dynamics simulations under electric field with millions of atoms.

Disordered structures and interface structures often require large cells to be representative of real systems. This makes their analysis challenging. By developing appropriate descriptors and unsupervised learning methods such as two-step locality preserving projections (TS-LPP), it becomes possible to detect subtle structural differences within a sample, without the need of prior knowledge. We are actively optimizing these methods to make it applicable to large complex systems.