Integrating Machine Learning and First-Principles Methods in Quantum Chemistry: Toward Accurate and Efficient Modeling of Solvation, Binding, and Catalysis

This dissertation addresses the challenge of improving the accuracy-efficiency balance in quantum chemical simulations of complex molecular systems. It spans three major themes: efficient computation of solvation free energies, accurate ranking of protein-ligand binding affinities, and surface catalytic behavior of transition metal phosphides. For solvation energy calculations, we investigated the application of delta-machine learning (Δ-ML) models. While conventional high-level methods are accurate, their computational cost is prohibitive for larger systems. Δ-ML bridges this gap by learning correction terms on top of lower-level quantum mechanical results. Our benchmarks across various solvent environments confirm that Δ-ML provides reliable predictions with a fraction of the computational expense, maintaining both generalization capability and chemical accuracy. For the ranking of protein-ligand binding affinities, this study uses two subsets of CDK2 protein complexes-CDK2-12 (PDB code: 2VT-) and CDK2-6 (PDB code: 4FK-)-as benchmark datasets. Several computational approaches were evaluated for their ability to produce accurate ranking correlations, including the Generalized Many-Body Expansion (GMBE), semi-empirical quantum methods, the D3-ML approach, and the Sfcnn deep learning model. Linear correlation analyses between predicted and experimental binding affinities were conducted for each method across both datasets, highlighting their relative strengths, stability, and applicability to systems of varying molecular complexity. These findings provide data-driven guidance for energy ranking in drug discovery and molecular screening workflows. In the catalytic domain, we employ periodic boundary conditions and first-principles calculations to investigate the surface properties of transition metal phosphides. By constructing representative catalytic surfaces and analyzing adsorption energies and reaction barriers, we uncover the active sites and surface-specific reaction mechanisms. These insights contribute to a deeper understanding of structure-function relationships in heterogeneous catalysis. In conclusion, this work presents a cohesive framework combining machine learning, quantum chemistry, and materials modeling to tackle complex molecular and catalytic systems. The methodologies developed herein provide scalable and accurate tools for predicting thermodynamic and kinetic properties in realistic chemical environments.