Highly Intelligent Forward Design of Metamaterials Empowered by Circuit‐Physics‐Driven Deep Learning

Abstract The field of advanced metamaterial design has witnessed the great progress of artificial intelligence (AI) for science. Typically, the widely used deep learning is a purely data‐driven approach. However, it is often assumed as a black box performing interpolation among training data in a fuzzy way, which suffers from poor generalization outside the training domain of critical physical parameters. Inspired by physics‐guided deep learning, a circuit‐theory‐informed neural network (CTINN) in order to strengthen the generalization of metamaterial design is proposed. A series of plasmonic stack metamaterials (PSMs) are taken as learning paradigms of CTINN. Compared to conventional deep learning, the scheme decreases one‐order lower test loss of spectral prediction beyond the structure training span, and it possesses an extraordinary function to predict spectra of the extrapolated wavelength range. Due to the physics‐informed mechanism, the CTINN only adopts 10% training samples of the pure data‐driven counterpart, while it reduces >50% test loss. Moreover, the CTINN design is used to guide the PSM experiments and demonstrate its great potential for metamaterial development. This work introduces the theory of equivalent circuits into the neural network, empowers physics supervision to AI, and accomplishes the smart and powerful design of PSMs.