Predictive modeling of composite fracture toughness using machine learning.

Area of Science:

• Materials Science
• Computational Materials Science
• Biomaterials Engineering

Background:

• Predicting the fracture toughness (K1c) of ion-releasing resin-based composites is crucial for dental material development.
• Understanding the influence of filler composition (barium glass, dicalcium phosphate dihydrate) and degree of conversion on K1c is essential.
• Machine learning offers a promising approach to model complex material property relationships.

Purpose of the Study:

• To apply machine learning models for predicting the fracture toughness (K1c) of experimental ion-releasing resin-based composites.
• To evaluate the impact of varying dataset sizes on the predictive performance and reliability of different machine learning models.
• To identify the most effective machine learning algorithms for this specific material property prediction task.

Main Methods:

• 234 K1c values from 21 composite formulations with varying barium glass and dicalcium phosphate dihydrate (DCPD) ratios were analyzed.
• Degree of conversion (DC) was included as a predictive variable.
• Four machine learning models (Penalized Regression, Random Forest, XGBoost, Neural Networks) were trained and tested on datasets of varying sizes (n=164, 88, 50 for training; n=70 for testing).

Main Results:

• XGBoost and Random Forest demonstrated the highest predictive performance (RMSE 0.120 and 0.123, respectively) and robustness across different training set sizes.
• Penalized Regression showed limited ability to capture complex interactions (RMSE 0.208).
• Neural Networks exhibited high sensitivity to reduced datasets, with significantly lower accuracy when trained on smaller data (RMSE 0.728 with n=50).

Conclusions:

• Ensemble-based machine learning models, specifically Random Forest and XGBoost, are effective for predicting composite K1c.
• These models provide reliable predictions even with relatively small training datasets.
• While acceptable predictions are possible with limited data, larger datasets enhance model reliability and generalizability for ion-releasing resin-based composites.