Deep-learning-based polarization-dependent switching metasurface in dual-band for optical communication

To address the critical limitations of conventional band-switching technologies - such as their slow speed, high energy consumption, and mechanical instability - this research introduces a novel deep-learning-driven framework for the intelligent inverse design of polarization-multiplexed metasurfaces. This approach represents a paradigm shift from traditional methods by enabling single-step, computational discovery of metasurface designs that directly encode two distinct optical functions within a single flat device. At the heart of our framework is a custom-designed deep neural network that seamlessly integrates parallel convolutional layers for robust feature extraction with cascaded regression modules for high-precision prediction. This hybrid architecture allows us to engineer sub-wavelength meta-atoms to achieve desired optical responses rigorously. As a groundbreaking demonstration, we designed and optimized a metasurface that achieves dynamic band switching solely through polarization modulation: it generates a targeted transmission peak in the O-band (1,260-1,360 nm) under y-polarization and an independent peak in the C-band (1,530-1,565 nm) under x-polarization. This mechanism eliminates the need for moving parts. The resulting device exhibits a switching efficiency orders of magnitude greater than its mechanical counterparts, while simultaneously offering enhanced stability, lower power consumption, and inherent adaptability for reconfigurable optical networks. Our work not only validates a specific device but also establishes a robust and generalizable design paradigm, underscoring the transformative potential of uniting deep learning with metasurfaces to achieve ultra-fast, intelligent, and efficient photonic systems for next-generation optical communications.