Robustness under parameter and problem domain alterations of Bayesian optimization methods for chemical reactions.

Area of Science:

• Chemistry
• Materials Science
• Computational Science

Background:

• Chemical reaction optimization and scope search aim to maximize product yield.
• Vast search spaces necessitate mathematical methods for efficient condition discovery.
• Bayesian optimization builds function approximations from limited data.

Purpose of the Study:

• Evaluate the robustness of the Experimental Design for Bayesian Optimization (EDBO) optimizer.
• Investigate the impact of altering acquisition functions and batch sizes on EDBO performance.
• Assess EDBO's applicability to new domains like molecular power conversion efficiency.

Main Methods:

• Modified the acquisition function and batch size of the EDBO optimizer.
• Applied EDBO to existing chemical reaction yield datasets.
• Tested EDBO on molecular power conversion efficiency data for photovoltaic cells.

Main Results:

• EDBO demonstrated broad robustness across tested modifications and datasets.
• Thompson Sampling acquisition function performed competitively against Expected Improvement.
• Observed performance concerns for EDBO in "incomplete" input domains.

Conclusions:

• EDBO is a versatile and robust optimization tool for chemical and materials science.
• Thompson Sampling presents a viable, computationally efficient alternative acquisition function.
• Further research is needed to address EDBO's limitations with incomplete input domains.