Explainable AI for Material Property Prediction Based on Energy Cloud: A Shapley-Driven Approach

Area of Science:

• Materials Science
• Machine Learning
• Ceramics Engineering

Background:

• Machine learning models in materials science often act as "black boxes", hindering interpretability.
• Predicting properties of lead zirconate titanate (PZT) ceramics is crucial for materials discovery.
• Assessing model performance and understanding variable contributions are essential.

Purpose of the Study:

• To predict the dielectric constant of PZT ceramics using the TabNet deep learning framework.
• To enhance the interpretability of machine learning models in materials science.
• To identify key components and process parameters influencing PZT dielectric properties.

Main Methods:

• Utilized the TabNet deep learning framework for property prediction.
• Employed Shapley Additive Explanations (SHAP) for model interpretability.
• Implemented various cross-validation techniques to ensure model reliability.

Main Results:

• TabNet significantly outperformed traditional machine learning models, achieving MSE of 0.047 and MAE of 0.042.
• SHAP analysis identified d33, tangent loss, and chemical formula as key predictors.
• Process time was found to be less influential on the dielectric constant prediction.

Conclusions:

• The TabNet model offers a transparent and accurate approach to predicting PZT dielectric properties.
• SHAP analysis provides valuable insights into the relationship between material components, processes, and dielectric behavior.
• This research advances materials discovery and predictive modeling for PZT ceramics.