Operator-induced structural variable selection for identifying materials genes.

Area of Science:

• Materials Informatics
• Computational Materials Science
• Catalysis

Background:

• Materials informatics faces challenges in identifying meaningful descriptors due to high-dimensional, correlated predictor spaces and limited sample sizes.
• Current methods struggle with the vast number of potential material descriptors derived from primary features and algebraic compositions.

Purpose of the Study:

• To develop a new dimension reduction method for materials informatics that overcomes the limitations of existing approaches.
• To enable efficient and accurate selection of physicochemically meaningful material descriptors (materials genes).

Main Methods:

• Formulation of the problem as variable selection with operator-induced structure (OIS).
• Iterative nonparametric variable selection for dimension reduction, utilizing the geometry embedded in OIS.
• Employment of a Bayesian Additive Regression Trees (BART)-based method for nonparametric module selection.

Main Results:

• The proposed method achieves dimension reduction orders of magnitude faster than existing techniques with improved accuracy.
• Demonstrated robust performance even with extremely high input dimensions, surpassing current methods.
• Identified key physical descriptors for single-atom catalysis, explaining metal-support binding energy and offering insights into sintering prevention.

Conclusions:

• The OIS framework and BART-based selection provide an effective and efficient approach for materials informatics.
• The method facilitates the discovery of interpretable physical descriptors, guiding catalyst design and material stability.
• This work advances the field by enabling analysis of previously intractable high-dimensional materials data.