Composilator_SupportingData

Accurate prediction of composite stiffness before fabrication remains difficult because real polymer composites deviate from classical mixture rules due to morphology, dispersion, and interphase effects. Micromechanical models provide physical bounds but cannot capture systematic morphology-induced bias, while unconstrained machine-learning models often fail to generalize across composite families.
This dataset underpins ComposiLator, a hybrid AI–mechanistic framework that refines the Voigt–Reuss–Hill midpoint using a physically constrained machine-learning correction factor informed by structured descriptors.
The curated corpus spans structurally independent polymer composite systems across diverse morphologies and reinforcement levels, enabling cross-family validation.
Using this dataset, ComposiLator reduced normalized prediction error by 86.5% (MAE = 0.784 GPa), achieved positive cross-system correlation (R² = 0.645), and outperformed Halpin–Tsai formulations within physically reasonable parameter ranges.