Adaptive representation of molecules and materials in Bayesian optimization.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Bayesian optimization (BO) is vital for automated materials discovery but relies heavily on effective molecular representations.
• Fixed representations, chosen by experts or data-driven methods, can be suboptimal or biased, especially for novel tasks with limited data.
• The completeness and compactness of feature vectors significantly impact BO efficiency.

Purpose of the Study:

• To introduce a Feature Adaptive Bayesian Optimization (FABO) framework that integrates dynamic feature selection into the BO process.
• To enable BO to adapt material representations throughout optimization cycles, improving efficiency and accuracy.
• To address limitations of fixed representations in novel molecular optimization tasks.

Main Methods:

• Developed FABO, a framework integrating feature selection with Gaussian processes within the Bayesian optimization loop.
• Dynamically adapted material representations (feature vectors) across multiple optimization cycles.
• Evaluated FABO on molecular optimization tasks, including the discovery of metal-organic frameworks (MOFs).

Main Results:

• FABO demonstrated superior performance compared to random search and methods using pre-defined feature spaces.
• The adaptive representation approach successfully identified effective features for distinct MOF discovery tasks with varying property distributions.
• FABO aligned with human chemical intuition for known tasks and proved robust for novel optimization challenges.

Conclusions:

• FABO offers a robust approach for navigating complex materials search spaces in automated discovery.
• Dynamic feature adaptation is crucial for optimizing Bayesian optimization performance, especially when prior knowledge is limited.
• The framework enhances the efficiency and reliability of automated materials discovery campaigns.