Berlin 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 23: Modeling and Simulation of Soft Matter III

CPP 23.3: Talk

Wednesday, March 20, 2024, 10:00–10:15, H 0107

Vibrational Spectroscopy from Machine Learning Molecular Dynamics by Accurately Representing the Atomic Polar Tensor — •Philipp Schienbein — Department of Physics, Imperial College London, London, SW72AZ, United Kingdom — Department of Physics and Astronomy, University College London, London, WC1E 6BT, United Kingdom

Vibrational spectroscopy is a key technique to elucidate microscopic structure and dynamics. Without the aid of theoretical approaches, it is, however, often difficult to understand such spectra at a microscopic level. Ab initio molecular dynamics has repeatedly proved to be suitable for this purpose, but the computational cost can be daunting; in particular when electronic structure methods beyond GGA DFT are required. Here, a new route to calculate accurate IR spectra from machine learning molecular dynamics is presented, utilizing the atomic polar tensor. The latter can be trained a posteriori on existing molecular dynamics simulations using the E(3)-equivariant neural network e3nn and is a most fundamental physical observable. The introduced methodology is therefore general and transferable to a broad range of systems. Besides benchmarking the method against explicit ab initio molecular dynamics, I will also present applications utilizing a atomic polar tensor neural network at the hybrid DFT level. These demonstrate that it has the potential to significantly contribute toward novel physical findings, especially where large-scale molecular dynamics simulations or expensive electronic structure calculations are required.

Keywords: machine learning; vibrational spectroscopy; molecular dynamics