Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MO: Fachverband Molekülphysik
MO 1: Ultrafast Structural Dynamics
MO 1.3: Vortrag
Montag, 2. März 2026, 12:15–12:30, P 105
Interpolating Grid Potential Energy Surfaces with X-MACE — •Paul Idzko1, Daniel Bitterlich2, Julia Westermayr2, and Daniel Keefer1 — 1Max-Planck-Institute for Polymer Research, Mainz, Germany — 2Wilhelm-Ostwald-Institute for Physical and Theoretical Chemistry - Leipzig University, Leipzig, Germany
The dynamics of electronically excited molecules is most accurately calculated by solving the time dependent Schrödinger equation (TDSE) on a grid. This gives access to the full nuclear wavefunction. A major challenge is to ensure high accuracy of the underlying electronic potential energy surfaces (PES). After performing quantum chemistry calculations on a sparse grid, interpolation to a much finer grid - usable in the dynamics simulations - becomes necessary. This especially applies to Conical Intersections, which exhibit cusps that are hard to interpolate via mathematical methods. X-MACE was presented as a deep learning architecture built upon the popular MACE program suite, enabling the learning of several electronic states, as well as the nonadiabatic coupling elements between the states. Here, X-MACE will be used to interpolate between the geometries calculated with quantum chemistry in an active learning approach. We investigate the influence of the chemical interpolation with X-MACE in contrast to mathematical interpolation (e.g. splines, polynomials...) on the quantum dynamic simulations for the system of 2,5-dichlorofuran.
Keywords: Potential Energy Surface; Machine Learning; MACE; Excited State; Quantum Dynamics