Modeling the dielectric constants of crystals using machine learning.

Area of Science:

• Materials Science
• Condensed Matter Physics
• Computational Chemistry

Background:

• Relative permittivity is key to crystal electromagnetic response, linking chemical bonding to macroscopic properties.
• Existing models (analytical, numerical, statistical) have limitations in scope or interpretability.
• Machine learning (ML) models offer predictive power but often lack physical explanations.

Purpose of the Study:

• To develop interpretable ML models for predicting optical dielectric constants in crystals.
• To combine supervised ML with physics-based analysis for enhanced understanding.
• To validate ML predictions against established physical models.

Main Methods:

• Trained support vector regression and deep neural networks on 1364 crystal dielectric constants.
• Utilized supervised machine learning (ML).
• Employed density functional perturbation theory and game theory for interpretability analysis (Shapley additive explanations).

Main Results:

• Achieved superior predictive power for optical dielectric constants compared to traditional methods.
• ML models recovered correlations described by Clausius-Mossotti and Penn models.
• Shapley additive explanations provided physical insights into the ML model's predictions.

Conclusions:

• The combined ML and interpretability approach accurately predicts crystal dielectric constants.
• This method enhances confidence in ML models by linking them to established physical principles.
• Offers a powerful tool for understanding and predicting dielectric behavior in crystalline materials.