Design and Optimization of Full-Stokes Hyperspectro-Polarimetric Encoding Metasurfaces Based on Conditional Multi-Task Deep Learning.

Metasurfaces, as two-dimensional metamaterials, offer an unprecedented avenue for multidimensional optical detection. However, existing full-Stokes hyperspectro-polarimetric encoding metasurface (FHPEM) designs rely on iterative trial-and-error processes and single-structure optimization, limiting design efficiency and broadband precision. Here, an end-to-end conditional multitask learning framework is proposed that integrates physical encoding constraints into the metasurface design process. By explicitly incorporating correlation constraints as physical conditions during network training, this framework establishes an on-demand metasurface design paradigm that directly generates structures with controlled encoding independence. Joint multitask learning further enables unified forward and inverse modeling across multiple metasurface configurations. Furthermore, a metasurface array screening strategy is developed by integrating forward prediction and correlation-conditioned inverse design networks. Using this strategy, metasurface arrays with sizes of 4 × 4, 6 × 6, 8 × 8, and 10 × 10 are designed, achieving average Relevances reductions of 19.7%, 34.6%, 26.9%, and 36.5%, respectively, compared with arrays obtained through manual selection or machine-learning-based methods. A 4 × 4 metasurface array is fabricated and experimentally validated, demonstrating full-Stokes spectro-polarimetric reconstruction over 400-900 nm with a spectral resolution of 4 nm. These results demonstrate the potential of the proposed approach for high-performance and highly integrated multidimensional optical systems.