AI-enabled structure-property prediction of woven thermoplastic composites from mesostructural RVE images

Woven thermoplastic composites are attractive for lightweight engineering applications, but their elastic response is strongly governed by mesoscale architectural features such as yarn width, yarn spacing, crimp, and matrix-rich regions. High-fidelity finite element homogenization can capture these structure–property relationships, but its computational cost limits its use for rapid materials design and large parametric exploration. This study develops an image-based deep learning surrogate for predicting the effective elastic properties of plain-weave carbon/polyphenylene sulfide composites from mesoscale representative volume element representations. A synthetic dataset of 999 woven composite RVEs with systematic yarn-geometry variations was generated using TexGen and simulated in Abaqus under periodic boundary conditions. Dual-view RVE images and the yarn fiber volume fraction were used as inputs to a multitask CNN–MLP model trained to predict domain fiber volume fraction, E 11 , E 22 , G 12 , and ν 12 . The model achieved high predictive accuracy for all outputs, with R 2 values above 0.98, while reducing the average evaluation time from 25.05 s per RVE in finite element simulation to 4.94 ms per RVE. Beyond reproducing FEM-based homogenized properties, the framework captures mesoscale geometry effects, including the influence of yarn compaction and crimp on stiffness and Poisson’s ratio. The proposed approach provides a rapid AI-enabled structure–property mapping tool for woven thermoplastic composites and offers a practical pathway toward accelerated composite materials design.