Multi-scale finite element analysis integrated with machine learning for efficient prediction of thermal conductivity in 3D orthogonal woven composites

3D orthogonal woven composites (3DOWCs) have attracted considerable research attention as highly reliable structural materials, primarily due to their unique spatial interwoven structure that exhibits excellent mechanical and thermal stability under harsh working conditions. However, the inherent complexity of their multiscale structure poses significant challenges for thermal conductivity prediction, with traditional methods relying heavily on extensive experiments and incurring high computational costs. To address this issue, this study proposes a multidimensional framework integrating the finite element method (FEM) and machine learning (ML) to replace conventional models for investigating 3DOWCs' effective thermal conductivity. 3DOWC models with various geometric parameters were constructed using Python scripts and TexGen, and a thermal conductivity dataset was obtained via multiscale FEM. Experimental validation using the flash method confirmed FEA reliability, after which combined finite element and experimental data trained ML models, comparing Kriging and artificial neural network (ANN) performance. Results show the Kriging model outperforms traditional approaches and ANN in computational efficiency and accuracy. Additionally, positive correlation between fiber volume fraction and thermal conductivity, and negative correlation with yarn spacing, were identified. This study presents an accurate, efficient prediction method to optimize 3DOWC design for enhanced thermal performance.