Abstract
Polymer composite properties are governed by multiple scales, yet heterogeneous physical descriptions and data representations across scales raise two challenges: what information to transfer and how to transfer it without a predefined unified mapping. We propose a generalizable multiscale information transmission framework (MITF), demonstrated on carbon fiber-reinforced liquid crystal polymer composites (CF-LCPs). MITF employs machine learning to identify and encode molecular information relevant to the target property, then uses a Generative Adversarial Network (GAN) to generate candidate molecules that retain this information as cross-scale carriers, constructing a composite design space. Material properties, structural parameters, and geometric information are integrated into a multimodal model linking heterogeneous cross-scale information to macroscopic properties, achieving a cross-validated R2 of 0.997. Inverse design yielded a target composite with an experimental modulus of 135.9 GPa. Thermo-mechanically coupled finite element simulations provided a qualitative assessment for the transmission, showing preferential initial damage in matrix and interfacial regions and interfacial debonding, consistent with established failure mechanisms. MITF enables effective cross-scale information transfer, providing a general methodology for multiscale intelligent design of polymer composites and other advanced materials with high-dimensional design spaces, heterogeneous representations, and limited data.