From microstructure to damage: a deep learning review for fiber-reinforced composite micromechanical response prediction

Résumé fourni par la source

Carbon fiber-reinforced polymers (CFRPs) have become indispensable in high-performance structural applications because of their high specific strength, stiffness, corrosion resistance and design flexibility; however, their mechanical behavior is governed by heterogeneous microstructures and manufacturing-induced defects that make accurate structure–property prediction difficult using conventional approaches. Traditional evaluation methods based on destructive experiments and multiscale finite element analysis (FEA) provide high fidelity but remain limited by time-intensive workflows, high computational cost, and simplifying assumptions regarding microstructural uniformity. These limitations have accelerated the adoption of machine learning, particularly deep learning, as an efficient surrogate framework for forward prediction of mechanical response directly from microstructural representations. Recent studies have shown that deep learning models can predict effective properties, full-field stress and strain responses, and damage evolution directly from image-based, feature-based, or hybrid representations of CFRP microstructures. However, existing reviews largely address AI in composites from a broad lifecycle perspective and do not systematically synthesize deep learning methods developed specifically for micromechanical structure–property prediction in CFRPs. This review fills that gap by classifying the literature according to microstructural input representation, physics integration level, and mechanical behavior regime (from linear elasticity to damage and failure), while also examining data generation strategies, computational efficiency, validation practice, and generalization capability to clarify current progress and identify priorities for robust, physically consistent, and industrially relevant predictive modeling.