Die DPG-Frühjahrstagung in Hannover musste abgesagt werden! Lesen Sie mehr ...

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

MO: Fachverband Molekülphysik

MO 17: Theory

MO 17.1: Vortrag

Donnerstag, 12. März 2020, 14:00–14:15, f142

Towards Machine-Learned Coordinate Grids for Wave Packet Dynamics — •Sebastian Reiter, Thomas Schnappinger, and Regina de Vivie-Riedle — Department of Chemistry, LMU Munich

The dynamics of ultrafast (photo)chemical processes are frequently studied quantum mechanically by propagating wave packets on a spatial grid of nuclear coordinates, thus solving the time-dependent Schrödinger equation. Here, dimensionality reduction is imperative for all but the smallest systems, as the number of grid points scales exponentially with the number of dimensions. This issue is commonly addressed by manually constructing a reduced-dimensional subspace that describes the process in question, for example by employing a few selected normal modes or a linear combination thereof as basis vectors. However, finding such a subspace can prove challenging as it requires a large amount of prior knowledge about the system.

We therefore present a semi-automatic technique to generate non-linear coordinate grids for use in quantum dynamics. It relies on a special type of artificial neural network, called autoencoder, which is capable of learning a low-dimensional representation of trajectory data. Starting from standard quantum chemical reaction path calculations, our software package is designed to assist the user in generating a suitable data set of molecular geometries, setting up and training the neural network and finally constructing the grid. We discuss the advantages of using non-linear over linear coordinate subspaces and present applications for our technique to quantum dynamics in both the ground state and excited states.