Machine learning approaches to optimization in chemical compound space

This thesis develops data-driven strategies for optimization in molecular descriptor-property space, a central challenge in chemistry with applications ranging from drug discovery to catalyst design. Molecular optimization is inherently difficult due to the combinatorial complexity of chemical systems and the high cost of property evaluation. These limiting factors demand the development of data-efficient methods for molecular property prediction and structural optimization. The thesis begins by developing classical machine learning (ML) approaches. I develop data-efficient strategies to accurately interpolate molecular properties across descriptor-property space, a key prerequisite for effective optimization. Central to this approach is the design of low-dimensional, physics-informed molecular descriptors. These descriptors help overcome limitations associated with conventional high-dimensional representations, which typically require large datasets for ML model training. Combined with Gaussian process regression (GPR) models employing compositional kernels selected using the Bayesian information criterion, this framework mitigates the curse of dimensionality and enables property prediction at chemical accuracy with small datasets. Building on this interpolation framework, I integrate GPR with Bayesian optimization (BO) to efficiently identify molecules with optimal properties. To complete the optimization pipeline, I introduce an inverse-mapping scheme that converts descriptor vectors proposed by BO into chemically valid molecular structures. I extend my study to more complex organometallic catalyst systems. A key bottleneck is generating diverse organometallic complexes with varied ligand environments and numerous isomers for chemical space exploration. To overcome this challenge, I develop an automated framework for the systematic generation and property evaluation of the complexes. The second part of the thesis explores the feasibility of using quantum machine learning (QML) based on analog neutral atom quantum computers for chemical space interpolation. I implement GPR models with quantum evolution kernels realized through programmable many-body dynamics in neutral atom arrays, extending QML from classification to regression tasks for molecular property prediction. Finally, I investigate quantum echo dynamics of cat states in a Kerr oscillator under parametric drive. I show that properly timed perturbations can partially reverse decoherence, even in the presence of dissipation that suppresses revival. This theoretical study provides insights into using echo as a tool for decoherence mitigation and quantum state characterization.