Regensburg 2025 – wissenschaftliches Programm

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

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 17: Modeling and Simulation of Soft Matter I

CPP 17.4: Vortrag

Dienstag, 18. März 2025, 10:30–10:45, H38

IR Spectroscopy and Electric Field Simulations Enabled by the Atomic Polar Tensor Neural Network — •Philipp Schienbein — Lehrstuhl für Theoretische Chemie II, Ruhr-Universität Bochum, 44780 Bochum, Germany — Research Center Chemical Sciences and Sustainability, Research Alliance Ruhr, 44780 Bochum, Germany

Vibrational spectroscopy is a vital technique for uncovering microscopic structure and dynamics of condensed phase systems. However, interpreting such spectra at the microscopic level often requires theoretical support. While ab initio molecular dynamics (AIMD) has proven effective for this purpose, its high computational cost -- especially when using electronic structure methods beyond GGA DFT -- can be prohibitive. We have recently introduced a machine learning molecular dynamics (MLMD) approach for accurately calculating IR spectra using the atomic polar tensor, a size-intensive physical observable that can uniquely be defined for each atom and is thus broadly applicable across diverse systems. Notably, the atomic polar tensor also rigorously enables MLMD simulations under external electric fields. In this work, I benchmark this method against explicit AIMD simulations demonstrating comparable accuracy with significantly reduced computational cost and present applications of the atomic polar tensor neural network. These examples highlight its potential to drive novel physical insights, particularly in scenarios requiring large-scale MD simulations or computationally intensive electronic structure methods.

Keywords: machine learning; molecular dynamics; vibrational spectroscopy; electric field simulations