Design of thermal cloaks with isotropic materials based on machine learning

Thermal manipulation has been widely researched due to its potentials in novel functions, such as cloaking, illusion and sensing. However, thermal manipulation is often realized by metamaterials which entails extreme material properties. Here, we propose a machine learning based thermal cloak consisting of a finite number of layers with isotropic materials. An artificial neural network is established to intelligently learn the relation between each layer's constitutive properties and the cloaking performances. Optimal material properties are retrieved so that heat flows can be directed to detour the cloaked object without any invasion, as if the object is not there. The designed cloak demonstrates both easiness to implement in applications and excellent performances in thermal invisibility, which are verified by simulations and experiments. The proposed method can be flexibly extended to other physical fields, like acoustics and electromagnetics, providing inspiration for metamaterials design in a wide range of communities.