Microstructural Factors Governing Fracture Toughness in Lithium Silicate Dental Glass-Ceramics Using Machine Learning

Abstract

The microstructure of lithium silicate-based dental glass-ceramics (LS-DGCs) is highly complex, consisting of multiple crystalline phases with diverse morphologies embedded within a residual glass matrix. This microstructure, which evolves through compositional changes and thermal treatment, strongly influences fracture resistance. Due to this complexity, identifying consistent microstructural factors within this commercial dataset that govern toughening becomes challenging. We established an experimentally grounded microstructure–property database for ten commercial LS-DGCs and utilized an Extremely Randomized Trees (ExtraTrees) regression model to capture the nonlinear relationships among numerous extracted microstructural descriptors. Indentation fracture toughness (K_c) was investigated alongside crack length (c) as a mathematically related consistency-check target. To reduce specimen-level data leakage, a fivefold GroupKFold validation strategy was used, with all images from the same physical specimen assigned to the same fold. Based on pooled out-of-fold predictions, the model achieved an R² of 0.671 for K_c. Fold-level performance varied across held-out specimen sets, indicating predictive utility within the present commercial dataset rather than uniform generalizability to every unseen specimen combination. Product-to-product and within-product correlation decomposition were used to examine whether descriptor associations extended beyond simple product identity. These analyses identified the LD aspect ratio and area fraction as candidate descriptors associated with K_c within the represented product families. Given the deterministic relationship between K_c and c, agreement in their SHAP rankings was interpreted as an internal consistency check rather than as independent mechanistic validation. This physics-guided framework helps distinguish the relative contributions of morphology-related descriptors and product-level compositional differences, providing complementary perspective on how microstructural evolution can lead to tougher LS-DGCs.