Advancing Computational Quantum Chemistry with Machine Learning

Abstract

This thesis focuses on advancing quantum chemistry using machine learning methods. For this purpose, a dataset for the lanthanoid elements is generated, conformer ranking is addressed leveraging novel machine learning architectures and the widely used extended tight-binding model is enhanced with automatic differentiation. <br/> The LnQM dataset, a comprehensive benchmark of 17269 mono-lanthanoid complexes optimized at PBE0-D4/def2-SVP level, enables systematic evaluations of quantum chemical and machine learning methods across the lanthanoid series. It features geometric, energetic, molecular and electronic properties at ωB97M-V/def2-SVPD level, granting insights into lanthanoid chemistry and highlighting limitations of current atomic charge models. <br/> The ConfRank ansatz improves conformer ranking through pairwise training of state-of-the-art machine learning models. Utilizing the DimeNet++ architecture, the accuracy of relative energy prediction on GMKTN55 conformational subsets is improved by 29% on average. Moreover, a considerable 100-fold computational speed-up compared to the currently used GFN2-xTB method is achieved using GPU infrastructure. <br/> The dxTB model, a PyTorch implementation of GFN-xTB, demonstrates a novel integration of quantum chemical algorithms into machine learning frameworks. It allows for differentiation of any input parameters to arbitrary order, achieving similar runtimes as the original Fortran implementation, which in turn lacks automatic differentiation. Moreover, parameter optimization can now be conducted using backpropagation, harnessing the extensive existing machine learning infrastructure, opening up possibilities to investigate new functional forms of internal xTB procedures and to develop individual, problem-specific GFN parametrizations. <br/> Together, these contributions chart new directions across different dimensions of computational research, ranging from data science aspects to model development. This thesis conduces to the ongoing advancement of machine learning in the domain of computational quantum chemistry and aims to offer a valuable contribution on the path to improved material sciences, healthcare and beyond.