Deep learning based design of thermal metadevices

Abstract

Thermal metadevices obtained from transformation optics have recently attracted wide attention due to their vast potential for thermal management. However, these devices require extreme material parameters that are difficult to achieve in large-scale applications. Here, we design a thermal concentrator using a machine learning method and demonstrate the thermal concentration performance of the designed device. We first define an architecture with a single isotropic material. Deep learning models based on artificial neural networks are implemented to retrieve design geometry parameters ensuring that the required spatially varying anisotropy is achieved. We implement the optimized architecture into a thermal concentrator, fabricate samples and experimentally demonstrate that the designed metamaterial can simultaneously concentrate the heat flux in its core and minimize perturbations to the external thermal field. Our approach paves new avenues for the design of thermal management devices and, more generally, enables feasible solutions for inverse heat manipulation problems.