Deep learning model for inverse design of interpenetrating phase composites with customizable thermal properties

A new deep learning model is developed for inverse design of interpenetrating phase composites (IPCs) with customizable thermal properties, including the components of the thermal conductivity (TC) and coefficient of thermal expansion (CTE) tensors. Hybrid triply periodic minimal surface (TPMS) structures are mathematically designed as the reinforcement phase to construct the new IPCs, which display orthotropic, tetragonal and cubic material symmetries. The effective TC and CTE tensors of the IPCs are determined using a finite element-based numerical homogenization method. A robust dataset containing six geometrical parameters and six thermal properties (including the three non-zero components of the TC and CTE tensors, respectively) is generated to establish the topology-property mapping. This is followed by the construction of a new tandem dual-network model, including a forward sub-model and an inverse sub-model, to predict the thermal properties and design the topologies of the hybrid TPMS-based IPCs. The hyperparameters are optimized for both the forward and inverse sub-models, which enables the newly proposed dual-network model to have excellent prediction and design capabilities. In addition, examples are provided to show that the thermal properties of the inversely designed IPCs match well with the targeted values within and beyond the original dataset. The current deep learning model provides a new tool for the forward prediction of thermal properties and inverse design of topologies of IPCs.