Multi-fidelity Bayesian optimization for targeted design space exploration.

Abstract

The aim of this thesis is to explore the application of Multi-Fidelity Bayesian Optimization (MFBO) for targeted design space exploration, specifically focusing on identifying high-performing Covalent Organic Frameworks (COFs) for methane storage. The primary challenge addressed in this work is the efficient navigation of a vast and complex design space with constrained computational resources.

To achieve this, a two-step MFBO approach was employed, combining high-fidelity simulations with low-fidelity approximations. This method significantly reduces the number of expensive high-fidelity evaluations required by leveraging correlations between different fidelity levels. The stopping criteria for the optimization process included a predefined iteration limit and an expected improvement threshold, ensuring an optimal balance between exploration, exploitation, and resource utilization.

The results demonstrate that the MFBO method outperforms traditional random sampling techniques, effectively identifying the top 100 COFs with the highest methane uptake. Kernel Density Estimates (KDE) and Principal Component Analysis (PCA) were used to visualize the optimization process, highlighting the efficiency of MFBO in focusing on high-potential regions of the design space.

The findings of this research offer a systematic and cost-effective approach to exploring large and complex design spaces. This methodology can be extended to other material classes and optimized further by integrating advanced machine learning techniques.