This study applied machine learning (ML) models to predict fracture toughness (K 1c ) of experimental ion-releasing resin-based composites and evaluated the impact of dataset size on prediction reliability. A total of 234 K 1c values from 21 composite formulations, containing varying ratios of barium glass and dicalcium phosphate dihydrate (DCPD; CaHPO₄·2H₂O), were analyzed. Total inorganic content ranged from 0 to 50 vol%, with “DCPD:glass” ratios adjusted in 10% increments. Degree of conversion (DC) was also used as a predictive variable. The dataset was split into three training sets (n = 164, 88, and 50) and a fixed test set (n = 70). Four ML models were evaluated: Penalized Regression (PR), Random Forest (RF), XGBoost, and Neural Networks (NN). XGBoost showed the highest predictive performance (RMSE = 0.120; MAE = 0.086; R² = 0.764), followed closely by RF (RMSE = 0.123; MAE = 0.093). Both maintained high accuracy across all training sizes, indicating robustness to data scarcity. PR (RMSE = 0.208; R² = 0.291) exhibited limited ability to model complex interactions, while NN showed high sensitivity to reduced datasets (RMSE = 0.728 with n = 50). The ensemble-based ML models Random Forest and Extreme Gradient Boosting were effective in predicting composite K1c, showing reliable results even with relatively small training sets. While acceptable predictions can be achieved with limited data, larger datasets improve model reliability and generalizability. • Machine learning (ML) models predicted fracture toughness of dental composites. • Experimentally derived dataset ensured robust and interpretable ML performance. • Ensemble models (XGBoost, Random Forest) achieved the highest prediction accuracy. • Approach may reduce redundant tests and accelerate material development.
Tonin et al. (Sun,) studied this question.