Improving Symbolic Regression for Predicting Materials Properties with Iterative Variable Selection.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Symbolic regression (SR) is valuable for deriving explicit structure-property relationships in materials.
• High-dimensional data and unclear key variables pose significant challenges for traditional SR methods.
• Existing SR techniques struggle with computational complexity when dealing with a large number of input features.

Purpose of the Study:

• To develop an automated approach for identifying important variables in complex materials science problems.
• To enhance the efficiency and accuracy of symbolic regression for materials discovery.
• To overcome the computational limitations of SR methods when faced with high-dimensional feature spaces.

Main Methods:

• Integration of symbolic regression (using SISSO) with iterative variable selection (VS).
• Development of a novel algorithm: VS-assisted SISSO (VS-SISSO).
• Utilized random search for the variable selection component within the algorithm.

Main Results:

• VS-SISSO effectively manages hundreds of input features, overcoming limitations of standalone SISSO.
• The algorithm demonstrates fast convergence towards optimal or near-optimal solutions for models with moderate complexity.
• Successfully applied to learn approximate equations for perovskite band gaps and single-atom alloy catalyst stability.

Conclusions:

• The proposed VS-SISSO approach significantly improves the accuracy and applicability of symbolic regression in materials science.
• Automated variable selection is crucial for enabling SR in complex, high-dimensional materials data.
• This method accelerates the discovery of explicit mathematical models for material properties.