A rapid feature selection method for catalyst design: Iterative Bayesian additive regression trees (iBART).

Area of Science:

• Computational Materials Science
• Machine Learning for Chemistry and Physics
• Data-driven Scientific Discovery

Background:

• Feature selection (FS) and feature engineering (FE) are crucial for developing predictive models in materials science.
• Traditional methods struggle with large, correlated feature spaces generated by recursive FE, leading to high computational demands.
• Existing FS methods face challenges with collinearity and computational complexity in large feature spaces.

Purpose of the Study:

• To develop a computationally efficient method for generating predictive descriptors by interleaving feature engineering and selection.
• To address the limitations of traditional FS approaches in handling highly correlated and massive feature spaces.
• To introduce iterative Bayesian additive regression trees (iBART) for progressive descriptor building and selection.

Main Methods:

• Developed iterative Bayesian additive regression trees (iBART), a novel method that interleaves feature engineering (FE) and feature selection (FS).
• iBART utilizes Bayesian additive regression trees (BART) for FS and FE with unary/binary operators in an iterative process.
• Compared iBART's performance against state-of-the-art FS methods (LASSO+l0, SIS, Bayesian FS) using catalysis metal-support interactions as a test case.

Main Results:

• iBART achieved performance comparable to existing state-of-the-art FS methods.
• The iBART method significantly reduced computational resources compared to traditional one-shot FE/FS approaches.
• iBART generated a maximum feature space of O(10^2), drastically smaller than the O(10^6) from one-shot methods.

Conclusions:

• The iBART method offers a computationally efficient alternative for developing data-driven descriptors in materials science.
• Interleaving FE and FS steps is an effective strategy to manage large feature spaces and reduce computational load.
• iBART demonstrates strong potential for accelerating the prediction of physical and chemical system properties.