Predicting and optimizing viscosity of dental resin composites with Gaussian process regression and Bayesian optimization.

OBJECTIVE: To develop a machine learning framework using Gaussian Process Regression (GPR) and Bayesian Optimization (BO) to predict and optimize the viscosity of resin composites at two shear rates (0.0106 and 74.4 s-1). METHODS: Fifty-four experimental resin composite formulations were prepared using two types of main fillers and one type of fumed silica filler under controlled conditions. Viscosity was measured using a rotational rheometer at 25 °C, and log-transformed values at 0.0106 and 74.4 s-1 were used as targets. The GPR models were trained with multiple kernel functions selected via 10-fold cross-validation (CV). Shapley Additive Explanations (SHAP) analysis was applied to interpret the feature contributions. BO with the Probability of Improvement acquisition function identified optimal formulations from 67,140 candidates. RESULTS: The predicted viscosity values at both shear rates were significantly correlated with the corresponding experimental data (p < 0.001, Pearson's correlation test). SHAP analysis revealed that fumed silica content had the greatest impact on viscosity at low shear rates, whereas main filler particle size and surface treatment were most influential at high shear rates. BO efficiently navigated the formulation space, achieving simultaneous optimization of V1 (0.0106 s-1) within the target range and reducing V2 (74.4 s-1) to the lowest quartile within only seven experimental iterations. SIGNIFICANCE: This study demonstrates the potential of GPR and BO for the data-driven design of dental resin composites. This approach enables rational optimization of handling properties, supporting clinical requirements for sculptability and extrudability. Future studies should expand the dataset and incorporate multi-objective optimization to balance viscosity with other critical properties such as mechanical strength and polymerization shrinkage.