Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
DY: Fachverband Dynamik und Statistische Physik
DY 14: Machine Learning in Dynamics and Statistical Physics II
DY 14.8: Talk
Monday, March 9, 2026, 17:00–17:15, HÜL/S186
Machine Learning for Electric-Field Driven Nuclear Dynamics in Solids and Liquids — •Elia Stocco1, Christian Carbogno2, and Mariana Rossi1 — 1MPI for the Structure and Dynamics of Matter, Hamburg, Germany — 2Fritz Haber Institute of the MPS, Berlin, Germany
Simulating the interaction of electric fields with matter is fundamental to study dielectric properties and their interplay with structural and vibrational degrees of freedom. Therefore, it is desirable to obtain a general method that is able to deal with static and time-dependent fields, that is scalable to large complex systems, and that retains ab initio accuracy. We describe a machine-learning molecular dynamics method within the electric dipole approximation that describes the coupling of insulating materials to diverse electric fields, spanning liquids, solids, and confined systems [1]. In particular, we also take into account the influence of the electric field on the lattice degrees of freedom. We train equivariant MACE models [2], using density-functional theory data to learn the potential energy and dipole surfaces, including the multi-valued nature of the polarization in periodic systems. The external forces on various degrees of freedom are obtained through automatic differentiation. We present calculations of the dielectric permittivity of water, the temperature and light-driven ferroelectric-paraelectric phase transition of LiNbO3, and simulations of piezoelectric systems. [1] Stocco, E., Carbogno, C. Rossi, M., Npj Comput. Mater. 11, 304 (2025). [2] Batatia, I., et al., Adv. Neural Inf. Proc. Sys. 35, 11423 (2022).
Keywords: ab initio molecular dynamics; equivariant neural networks; piezoelectricity; ferroelectricity; water