Efficient Inverse Design of Infrared Metasurfaces with Super-Unit via Image-Based Transitive Transfer Learning

Abstract Deep learning has emerged as a potent methodology for the inverse design of metasurfaces; however, its deployment in complex optical systems is often hindered by substantial computational demands and the scarcity of comprehensive data sets, particularly when addressing diverse design requirements. To address these challenges, we introduce a tandem network architecture that utilizes parameter-encoded images and develop an image-based transitive transfer learning (I–TTL) framework aimed at the efficient design of diversified infrared metasurfaces with superunit. By synergistically combining image feature extraction with transitive transfer learning, the proposed framework markedly reduces the necessity for extensive training data sets, accelerates model convergence, and improves predictive accuracy. The knowledge transfer is facilitated through a two-stage migration process, progressing sequentially from a source domain to an intermediate domain, and subsequently to the target domain. Experimental results demonstrate that both the intermediate and target domains achieve design accuracy comparable to that achieved by the independent training method, while using only 36% of the source-domain data volume. Moreover, when the number of training epochs and the design diversity are held constant, the proposed transitive transfer learning strategy via an intermediate domain delivers a more pronounced performance improvement than direct transfer learning. Utilizing this framework, we successfully designed three distinct infrared metasurfaces composed of varied material combinations. These metasurfaces exhibit low emissivity within atmospheric windows and maintain an average emissivity exceeding 0.6 outside these windows, thereby effectively reconciling stealth capabilities with radiative cooling requirements. This study presents a robust and generalizable approach for the efficient inverse design of a variety of plasmonic devices.