Predicting the Mechanical Behavior of Fiber-Reinforced Polymer Composites Through Data-Driven Multiscale Modeling

Abstract

Fiber-reinforced polymer composites exhibit mechanical behavior that emerges from coupled mechanisms across fiber, matrix, interface, and laminate scales. Computational homogenization can represent these mechanisms, but repeated fine-scale simulations are too expensive for design exploration, inverse identification, and uncertainty propagation. This manuscript develops a data-driven multiscale modeling framework for predicting effective elastic and failure-relevant response of fiber-reinforced polymer composites with quantified uncertainty. The framework combines representative volume element analyses, microstructure descriptors that encode fiber volume fraction, orientation, aspect ratio, spatial clustering, and interface-sensitive quantities, and a physics-constrained surrogate that maps descriptors to homogenized properties. A Gaussian-process component supplies nonparametric regression and predictive variance, while a physics-informed residual term enforces equilibrium-consistent trends and monotonic constraints derived from micromechanics. Multifidelity correction links inexpensive lower-fidelity estimates to higher-fidelity homogenization data, and Bayesian calibration treats uncertain model parameters and manufacturing-induced variability explicitly. The central contribution is an integrated formulation that connects microstructural descriptors, surrogate inference, physics constraints, multifidelity correction, and uncertainty propagation in one predictive chain. The analysis clarifies where data-driven acceleration is defensible, how descriptor choice affects extrapolation, and why calibration must separate parameter uncertainty from model discrepancy. The resulting framework provides a technically grounded route for predicting composite mechanical behavior without treating machine learning as a black-box substitute for mechanics.