Gradient boosted and statistical feature selection workflow for materials property predictions.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Data-driven approaches and machine learning (ML) are crucial for accelerating materials discovery.
• Current ML methods often lack interpretability and require computationally expensive regularization techniques.
• Identifying relevant features is key to improving ML model generalization and reducing computational cost.

Purpose of the Study:

• To develop and validate a feature selection workflow for ML in materials science.
• To enhance model interpretability and reduce computational expenses.
• To accelerate the discovery of new materials through efficient data analysis.

Main Methods:

• A recursive feature selection workflow leveraging gradient boosting and statistical analyses.
• Multicollinearity reduction using feature correlation and hierarchical clustering.
• Wrapper methods with greedy search for feature refinement.
• Bayesian optimization for ML model training.

Main Results:

• The workflow successfully identifies a relevant subset of features, maximizing predictive power.
• Minimal feature redundancy was achieved through multicollinearity reduction.
• ML models trained using the selected features, without regularization, achieved state-of-the-art performance.
• The workflow demonstrated generalizability across various material property predictions.

Conclusions:

• The proposed feature selection workflow effectively addresses ML interpretability and computational cost challenges in materials discovery.
• This approach enables the development of accurate and efficient ML models for predicting material properties.
• The workflow facilitates accelerated discovery of novel materials.