Explicitly Incorporating Prior Knowledge into Bayesian Optimization for Materials Design

Area of Science:

• Computational Chemistry
• Machine Learning
• Experimental Design

Background:

• Bayesian optimization (BO) is crucial for efficient experimental design.
• Incorporating prior knowledge into BO can accelerate optimization.
• Existing BO methods may not fully leverage domain-specific linear relationships.

Purpose of the Study:

• To develop a novel Bayesian optimization method that integrates prior knowledge of linear relationships between experimental conditions and target properties.
• To reduce the number of experiments required to identify optimal experimental parameters.
• To enhance the efficiency of optimization processes in scientific research.

Main Methods:

• Developed a Bayesian optimization (BO) approach using Gaussian process regression with a mean function incorporating prior knowledge of linear relationships.
• Modeled the relationship using multivariate polynomials with prior distributions on coefficients, reflecting vague knowledge and sign information.
• Employed Markov chain Monte Carlo (MCMC) sampling to derive posterior distributions and acquisition functions for selecting next experimental points.

Main Results:

• The proposed BO method successfully identified optimal experimental conditions with fewer trials compared to existing approaches.
• Performance was validated using mathematical simulation functions, a boiling-point dataset, and a pigment dataset.
• The method demonstrated superior efficiency when prior knowledge accurately reflected the underlying experimental-property relationship.

Conclusions:

• The novel Bayesian optimization method effectively utilizes prior knowledge of linear relationships to improve optimization efficiency.
• This approach offers a significant advantage in reducing experimental costs and time, particularly in chemistry and materials science.
• The study highlights the importance of incorporating domain expertise into machine learning models for scientific discovery.