A Comparative Empirical Study of Interpretable Regression and Clustering Methods for Quantifying Dental Composite Formulation-Performance Relationships

Abstract

Dental resin composites are the most widely used direct restorative materials, yet their formulation optimization remains predominantly empirical due to the complex, nonlinear interactions among resin monomers, fillers, and processing parameters. This paper presents a comparative empirical study evaluating six regression algorithms and three clustering methods for quantifying formulation-performance relationships in dental composites. Using a curated dataset of 233 dental composites spanning 17 formulation attributes and 7 mechanical and physical performance indicators, we benchmark Ridge Regression, Lasso, Random Forest, XGBoost, Support Vector Regression, and Gaussian Process Regression under nested five-fold cross-validation. Clustering analysis via K-means, DBSCAN, and agglomerative hierarchical clustering identifies distinct formulation archetypes with characteristic performance profiles. SHAP-based interpretability analysis reveals that filler loading, the BisGMA-to-TEGDMA monomer ratio, and degree of conversion collectively account for over 58% of total SHAP attribution across all target-parameter pairs. XGBoost achieves the highest average coefficient of determination (R2 = 0.891) across seven performance targets, while Gaussian Process Regression provides a principled uncertainty framework that may help guide future experimental design. The identified formulation parameter influence weights offer materials engineers a quantitative basis for prioritizing experimental variables, reducing trial-and-error costs in dental composite development. These findings support the Materials Genome Initiative's objective of accelerating materials screening through data-driven methods.