Intelligent design of thermal metamaterials based on deep learning assisted structure optimization

Thermal metamaterial has recently garnered increasing attention for manipulating heat flows and enabling diverse thermal functionalities. The conventional strategy to design thermal metamaterials is to use layered configurations and other structural layouts, whose design space is limited or the workload is heavy. A repeated design or optimization process is needed to retrieve desired architectures that occupy target thermal conductivity. Here, we propose an intelligent design approach (automatic and customized) as the alternative to conceiving thermal metamaterials by combining size optimization with deep learning. We show that this approach can generate desired mixture architectures that cover the full-parameter anisotropic space, with exceptional accuracy, speed, and efficiency. Transformation-induced thermal meta-devices can be constructed directly by assembling retrieved mixture architectures, regardless of the thermal functionalities. Different metadevices are designed and verified by simulations and experiments, demonstrating the feasibility and versatility of this strategy. The formulated method increases the design efficiency and holds better manufacturability, which holds promise for other physical fields, e.g., acoustics and mass diffusion.