Abstract:
The thesis comprises a series of publications in the emerging field of deep learning-based Variational Monte Carlo (DL-VMC), which offers a highly accurate ab-initio method for computing electronic properties of molecules by solving the Schrödinger equation. The time-independent, non-relativistic Schrödinger equation is an eigenvalue equation, whereas the thesis places particular emphasis on the smallest eigenvalue and its
corresponding eigenfunction. In DL-VMC, in an unsupervised manner, a neural network is optimized to represent the solution of the Schrödinger equation, also called the wave function. The method has recently demonstrated superior accuracy compared to other computational methodologies, albeit at a substantial computational cost due to the large number of trainable parameters in the neural network and the need for Monte Carlo integration to obtain a precise estimate of the eigenvalue.
The following thesis contributes to the recent improvements to enhance the method’s efficiency by introducing a novel technique for effectively learning wave functions for various molecules and their geometrical conformations in parallel. For the first time, we are able to optimize a single wave function model on a diverse range of molecules and transfer this pretrained neural network to previously unencountered molecules, yielding competitive results with only a few fine-tuning steps. Furthermore, this work aims to improve the understanding of specific design choices involved in constructing neural
network-based wave functions.

Zoom-Link:
https://univienna.zoom.us/j/68409697007?pwd=NEZiajBXcEtSL2JzOGVCT3lLMzBWUT09

Meeting-ID: 684 0969 7007
Kenncode: 776439