Bayesian Optimization for Materials Design with Mixed Quantitative and Qualitative Variables.

Area of Science:

• Computational Materials Engineering
• Materials Science
• Optimization

Background:

• Bayesian Optimization (BO) accelerates materials design but is limited to quantitative variables.
• Real-world materials design involves mixed qualitative and quantitative variables (composition, microstructure, processing).
• Existing methods for mixed variables use dummy variables, limiting correlation capture.

Purpose of the Study:

• To integrate a novel latent-variable (LV) approach for mixed-variable Gaussian process (GP) modeling into the BO framework.
• To address limitations of existing methods in capturing complex correlations between qualitative factors.
• To enhance the accuracy and interpretability of Bayesian Optimization for materials design.

Main Methods:

• Developed a latent-variable Gaussian process (LVGP) model for mixed-variable problems.
• Integrated LVGP into the Bayesian Optimization (BO) framework.
• Applied the approach to materials design examples, including solar cells and perovskites.

Main Results:

• The LVGP approach demonstrated superior modeling accuracy compared to existing methods.
• Mapped latent variables provided intuitive visualization and insights into qualitative factor effects.
• Successfully applied to optimize light absorption in solar cells and design perovskite materials.

Conclusions:

• The proposed LVGP-BO method effectively handles mixed-variable optimization problems in materials design.
• This approach offers a more flexible and accurate way to model qualitative factors.
• The methodology is generalizable to other complex design optimization problems with expensive simulations.