Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

O: Fachverband Oberflächenphysik

O 2: Focus Session: Frontiers of Electronic-Structure Theory – Advances in Time-Dependent and Nonequilibrium Ab Initio Methods I

O 2.3: Vortrag

Montag, 18. März 2024, 11:00–11:15, HE 101

First-principles light-driven molecular dynamics through equivariant neural networks — •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

Recent experiments have shown a rich phenomenology in solids, liquids and molecules when driven by ultra-fast THz pulses. However, simulation techniques that can describe the nuclear dynamics of these processes for all material classes, without relying on dimensionality reduction, and going beyond perturbation theory, are very challenging. Here we propose an ab initio molecular dynamics (MD) method within the electric dipole approximation that allows a single machine-learning model to describe the coupling at diverse field strengths and with time dependence. Our requirement is that the system remains near the electronic ground state. We train an equivariant differentiable neural network to learn the dipole of isolated and periodic systems. Atomic tensor derivatives w.r.t. the nuclear coordinates are obtained through autodifferentiation. We present applications of this machine-learning-assisted MD protocol on water and LiNbO3. We show full-dimensional ab initio simulations of the excitation of vibrational modes at an ultra-fast time scale and highlight the different nonlinearities of these dynamics.

Keywords: ab initio molecular dynamics; equivariant neural networks; light-matter coupling; light-driven system