Interpretable Bayesian optimization for catalyst discovery.

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Bayesian optimization (BO) is crucial for exploring complex material design spaces.
• Existing BO methods require prior knowledge of key physical parameters (features), which are often unknown in heterogeneous catalysis.
• The interplay of multiple physical processes complicates material property prediction.

Purpose of the Study:

• To introduce a novel Bayesian optimization framework, SARBO, for efficient materials discovery.
• To address the challenge of unknown key physical parameters in heterogeneous catalysis.
• To enable on-the-fly selection of relevant features using symbolic regression.

Main Methods:

• Developed the Sparse Adaptive Representation-based Bayesian Optimization (SARBO) framework.
• Integrated the sure independence screening and sparsifying operator (SISSO) symbolic-regression method for feature selection.
• Applied SARBO to simulated discovery of single- and dual-atom alloy surface sites for CO2 activation.

Main Results:

• SARBO successfully performs on-the-fly selection of key physical parameters, considering nonlinear relationships and interactions.
• The framework demonstrates efficient navigation of materials spaces.
• SARBO outperforms traditional feature-selection approaches in simulated catalyst discovery.

Conclusions:

• SARBO offers an effective solution for materials discovery when key features are initially unknown.
• The method enhances the efficiency of Bayesian optimization in complex chemical systems.
• SARBO facilitates the discovery of novel catalysts, such as those for CO2 activation.