Multiscale modeling of the effective thermal conductivity of 2D woven composites by mechanics of structure genome and neural networks

•An ultra-efficient, data-driven multiscale modeling approach is developed to predict the thermal conductivity of 2D woven composites.•Mechanics of structure genome theory is extended to predict the thermal conductivity of general textile composites.•Different weave patterns of 2D woven composites are converted to continuous input for the neural network models via one-hot encoding.•The developed data-driven multiscale models can accurately predict thermal conductivity of woven composites considering various microscale and mesoscale features. A data-driven multiscale modeling approach is developed to predict the effective thermal conductivity of two-dimensional (2D) woven composites. First, a two-step homogenization approach based on mechanics of structure genome (MSG) is developed to predict effective thermal conductivity. The accuracy and efficiency of the MSG model are compared with the representative volume element (RVE) model based on three-dimensional (3D) finite element analysis (FEA). Then, the simulation data is generated by the MSG model to train neural network models to predict the effective thermal conductivity of three 2D woven composites. The neural network models have mixed input features: continuous input (e.g., fiber volume fraction and yarn geometries) and discrete input (e.g., weave patterns). Moreover, the neural network models are trained with the normalized features to enable reusability. The results show that the developed data-driven models provide an ultra-efficient yet accurate approach for the thermal design and analysis of 2D woven composites.