Interpretable Machine Learning with Prediction Uncertainty Quantification for d₃₃ in (K_0.5Na_0.5) NbO₃-Based Lead-Free...

Area of Science:

• Materials Science
• Ceramics Engineering
• Computational Materials Science

Background:

• Discovering high-performance lead-free piezoelectric ceramics is challenging due to vast compositional spaces.
• Conventional machine learning (ML) models often lack interpretability, hindering material design.
• Lead-free piezoelectrics are crucial for sustainable electronic applications.

Purpose of the Study:

• To develop a physics-informed and interpretable ML framework for predicting piezoelectric properties.
• To integrate uncertainty quantification for reliable material predictions.
• To identify key factors governing the piezoelectric coefficient (d33) in (K0.5Na0.5)NbO3 (KNN)-based ceramics.

Main Methods:

• A curated dataset of 1113 KNN-based ceramic samples was utilized.
• Feature engineering decoupled A-site and B-site ionic contributions, reducing descriptors.
• Deep neural networks, Wide & Deep networks, residual networks, and Bayesian neural networks were employed.
• SHapley Additive exPlanations (SHAP) and Sure Independence Screening and Sparsifying Operator (SISSO) were used for interpretability.

Main Results:

• The ML framework achieved high prediction accuracy (R² ≈ 0.81) for the piezoelectric coefficient (d33).
• Predictive uncertainty from Bayesian neural networks quantified model confidence.
• A compact analytical descriptor identified sintering temperature, B-site electronic anisotropy, and A-site ionic displacement as key governing factors.
• The framework provided transparent design rules for lead-free piezoelectrics.

Conclusions:

• The proposed physics-informed ML framework accelerates the discovery of high-performance lead-free piezoelectric materials.
• Interpretability and uncertainty quantification are crucial for reliable ML-driven materials design.
• Understanding the interplay of sintering temperature, electronic anisotropy, and ionic displacement enables targeted optimization of KNN-based ceramics.