Accurate predictive model of band gap with selected important features based on explainable machine learning.

Area of Science:

• Materials Informatics
• Computational Materials Science
• Machine Learning

Background:

• Nonlinear machine learning models excel at predicting material properties but often lack interpretability.
• Black-box models may include irrelevant or detrimental features, impacting performance and trustworthiness.

Purpose of the Study:

• To apply explainable machine learning (XML) techniques to identify crucial features for predicting material band gaps.
• To develop a framework for constructing simplified, accurate, and interpretable predictive models using feature importance.

Main Methods:

• Employed permutation feature importance and SHapley Additive exPlanations (SHAP) on a support vector regression model.
• Utilized 18 input features to predict GW-level band gaps.
• Proposed a framework for reduced-feature model construction guided by XML insights, emphasizing the removal of highly correlated features ( > 0.8).

Main Results:

• An XML-guided model with the top five features achieved comparable accuracy to the original 18-feature model on in-domain data (0.254 vs. 0.247 eV).
• The reduced-feature model demonstrated improved generalization and lower prediction errors on out-of-domain data (0.348 vs. 0.460 eV).
• Eliminating strongly correlated features prior to XML analysis is crucial for accurate feature importance assessment.

Conclusions:

• XML techniques effectively clarify feature roles in machine learning models for materials informatics.
• Simplified models derived through XML offer comparable accuracy, improved generalization, and enhanced trustworthiness.
• This approach reduces computational expenses associated with feature selection and acquisition in materials discovery.