Polarimetry plays a pivotal role in diverse fields ranging from astronomy to biomedical imaging. However, the widespread adoption of full-Stokes polarimeters is constrained by the bulkiness of traditional optical components (e.g., wave plates) and the fabrication complexity of emerging metasurfaces, which typically require costly lithography and rigorous pixel-to-structure alignment. Here, we demonstrate a cost-effective, single-shot full-Stokes polarimeter based on a disordered twisted nanohole array (DTNA) metasurface, fabricated via scalable microsphere lithography. By leveraging the intrinsic long-range disorder and moiré-induced anisotropy, the device generates spatially diverse optical responses that robustly encode arbitrary polarization states. To decode this information without complex physical calibration, we develop a deep learning framework that establishes an end-to-end mapping between the spatially encoded intensity distribution and the full Stokes vector. The system achieves high-precision reconstruction with normalized mean squared errors (MSEs) of 0.047%, 0.16%, and 0.032% for S 1 , S 2 , and S 3 , respectively. Furthermore, through a systematic investigation of training data distribution, dataset volume, and super-pixel size, we identify the physical origin of this performance: the synergistic interaction of microscopic optical disorder (initial-phase, linear dichroism (LD), and circular dichroism (CD)) creates a high-dimensional feature space, where detection accuracy improves with increased disorder until reaching an information saturation point. This work not only offers a robust, alignment-free paradigm for high-performance integrated polarimetry but also provides fundamental insights into the role of disorder in computational optical sensing.
Wu et al. (Tue,) studied this question.