Interpretable Machine Learning for the Design of (K, Na)NbO₃-Based Piezoceramics Using Combinatorial and...

Area of Science:

• Materials Science
• Ceramics Engineering
• Computational Materials Science

Background:

• Potassium sodium niobate (KNN)-based ceramics exhibit promising piezoelectric properties, crucial for sensor and actuator applications.
• Enhancing KNN piezoelectricity via chemical modification is effective but lacks efficient screening methods for optimal dopant selection.
• Machine learning (ML) offers potential for predicting piezoelectric performance but often lacks interpretability.

Purpose of the Study:

• To develop an interpretable framework for identifying critical features influencing piezoelectric constant (d33) in KNN ceramics using ML.
• To quantitatively analyze the impact of material features on d33 predictions, uncovering nonlinear and interactive effects.
• To create combinatorial descriptors for improved ML model performance and interpretability in KNN material design.

Main Methods:

• Utilized Shapley Additive exPlanations (SHAP) to analyze feature importance and interactions in ML models predicting d33.
• Employed Sure Independence Screening and Sparsifying Operator (SISSO) to generate combinatorial descriptors from key features.
• Trained and evaluated ML models using individual features and novel combinatorial descriptors, assessing performance metrics like MAE and R^2.

Main Results:

• Identified critical features and their tipping points (e.g., sintering temperature) that significantly influence KNN piezoelectricity (d33).
• Developed combinatorial descriptors that exhibit approximate linearity with d33, enabling transparent linear models.
• Achieved high performance with ML models incorporating combinatorial descriptors, reaching MAE < 30 pC/N and R^2 > 0.9.

Conclusions:

• The developed interpretable framework effectively guides the rational design of KNN-based ceramics for enhanced piezoelectric performance.
• Feature engineering through combinatorial descriptors significantly improves both the interpretability and predictive accuracy of ML models.
• This feature-centric approach provides valuable insights into the complex mechanisms governing piezoelectric enhancement in KNN materials.