Bayesian optimisation in chemical problems
File(s)
Author(s)
Wu, Yifan
Type
Thesis
Abstract
Materials optimization plays a crucial role in accelerating the development of high-performance photovoltaic devices and other functional materials. Computational approaches, particularly Bayesian Optimization (BO), offer a promising avenue to efficiently navigate vast and complex chemical spaces, significantly reducing the time and cost associated with traditional experimental discovery.
In the first results chapter of this thesis (Chapter 4), the performance of various optimization strategies—including gradient descent, simulated annealing, and Bayesian Optimization—is evaluated on benchmark functions, demonstrating BO’s superior efficiency and reliability. Building on this, BO is applied to optimize photovoltaic device configurations to maximize solar cell efficiency. Chapter 5 introduces a optimization framework that integrates parallel Bayesian Optimization with Hyperband and transfer learning. This hybrid approach accelerate convergence and improve solution quality in multi-junction solar cell optimization. Domain-informed partitioning and adaptive evaluation further enhance the robustness of the optimization workflow. In Chapter 6, dimensionality reduction techniques are employed to simplify the high-dimensional chemical search spaces, making optimization more tractable and efficient. Principal Component Analysis (PCA) and related methods are shown to preserve essential chemical information, enabling Bayesian Optimization to perform effectively even in noisy or sparse data regimes. Chapter 7 expands upon molecular and material representations by exploring molecular fingerprints and machine learning force fields that generate vectorized descriptors of chemical structures. These improved representations enrich the target search space, facilitating more informed and efficient optimization of materials properties.
Together, the results presented in this thesis demonstrate how the integration of advanced optimization algorithms, dimensionality reduction, parallel computation, and informed material representations can drive forward the discovery and design of next-generation photovoltaic materials. Beyond solar cells, these methodologies contribute broadly to the field of materials informatics, providing a foundation for accelerated computational materials design.
In the first results chapter of this thesis (Chapter 4), the performance of various optimization strategies—including gradient descent, simulated annealing, and Bayesian Optimization—is evaluated on benchmark functions, demonstrating BO’s superior efficiency and reliability. Building on this, BO is applied to optimize photovoltaic device configurations to maximize solar cell efficiency. Chapter 5 introduces a optimization framework that integrates parallel Bayesian Optimization with Hyperband and transfer learning. This hybrid approach accelerate convergence and improve solution quality in multi-junction solar cell optimization. Domain-informed partitioning and adaptive evaluation further enhance the robustness of the optimization workflow. In Chapter 6, dimensionality reduction techniques are employed to simplify the high-dimensional chemical search spaces, making optimization more tractable and efficient. Principal Component Analysis (PCA) and related methods are shown to preserve essential chemical information, enabling Bayesian Optimization to perform effectively even in noisy or sparse data regimes. Chapter 7 expands upon molecular and material representations by exploring molecular fingerprints and machine learning force fields that generate vectorized descriptors of chemical structures. These improved representations enrich the target search space, facilitating more informed and efficient optimization of materials properties.
Together, the results presented in this thesis demonstrate how the integration of advanced optimization algorithms, dimensionality reduction, parallel computation, and informed material representations can drive forward the discovery and design of next-generation photovoltaic materials. Beyond solar cells, these methodologies contribute broadly to the field of materials informatics, providing a foundation for accelerated computational materials design.
Version
Open Access
Date Issued
2025-09-30
Date Awarded
2026-03-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Walsh, Aron
Ganose, Alexander
Publisher Department
Department of Materials
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)