Thermal Conductivity Prediction of Woven Composites Using a Large Language Model-Integrated Neural Network Framework

This paper presents a hybrid physics and machine learning framework for predicting the thermal conductivities of woven composites. A two-step homogenization procedure based on mechanics of structure genome (MSG), a top-down multiscale modeling approach, is used to compute effective thermal conductivities of woven composites. First, fiber and matrix at the microscale are homogenized to obtain effective thermal conductivity of yarn for subsequent mesoscale analysis. Then, the mesoscale model is homogenized to compute the effective thermal conductivities of woven composites, accounting for yarns, matrix, and different weave patterns. The resulting dataset is used to train an artificial neural network for rapid prediction of in-plane and out-of-plane thermal conductivities across a wide design space. The core contribution is the integration of this model with an AI-powered composite expert system, CompositesAI, which allows users to query designs and retrieve predictions conversationally without deep knowledge in the theory or the code itself. The framework also supports dynamic retraining as new physics-based data are generated, ensuring continuous model improvement. This approach combines multiscale modeling accuracy with AI-driven accessibility, enabling efficient and scalable design exploration in composite materials and structures.