Eye accommodation-inspired neuro-metasurface focusing

Focusing light has been a significant challenge for centuries. While conventional lenses are widely used, their bulkiness limits miniaturization and integration. Metasurfaces, composed of subwavelength scatterers, offer a solution for creating compact metalenses. However, most metalenses operate only in predefined environments, lacking the adaptability needed for real-world applications. Artificial intelligence, particularly deep learning, has shown promise in creating adaptive optics, but it relies heavily on extensive training data and struggles with rapidly changing environments. The human eye, with its ability to dynamically adjust its lens for focusing, serves as a compelling biological model for creating an intelligent, adaptive focusing system. This study leverages the principles of eye accommodation to develop a neuro-metasurface focusing system that exhibits self-adaptability and rapid response to complex electromagnetic environments.

Extensive research focuses on metalenses using various metasurface structures to achieve broadband, achromatic performance, and high efficiency. However, most designs lack adaptability to changing environments. While deep learning has shown potential for adaptive optics, its dependence on large training datasets and vulnerability to dynamic environments limits its practical application. The human eye's accommodation mechanism, which involves the ciliary muscles adjusting the lens shape for different focal distances, is a highly efficient and adaptive biological system, inspiring researchers to explore similar self-regulating mechanisms for artificial systems.

This work introduces the concept of Supervised Evolving Learning (SEL) to design an adaptive focusing neuro-metasurface (SELAF). The SEL algorithm involves two cycles: an operation cycle and an evolution cycle. In the operation cycle, the system interacts with the environment using a neuro-metasurface that adjusts its phase based on incident wave information collected by an array probe. A steering network predicts the required phase compensation, and a controller adjusts the metasurface voltage accordingly. The evolution cycle updates the steering network using a continuous optimization algorithm (Adam) based on data collected during the operation cycle. The neuro-metasurface consists of unit cells with embedded PIN diodes that allow for 180° phase reversal by adjusting the applied voltage. A convolutional neural network (CNN)-based focus steering network (FSN) maps detected electric fields to phase compensation values. A focus controller, employing the Adam algorithm, determines the voltage adjustments to achieve focusing. The system's performance was evaluated through simulations and experiments involving various scenarios: single-source incidence, dual-source incidence, and a scattering scenario. Focusing efficiency and the number of iterations were used to assess the system's performance.

The SELAF system demonstrated effective adaptive focusing in diverse EM environments. Experiments showed successful focusing with a single source, two sources, and an obstacle in the path. Focusing efficiencies ranged from 39% to 45%, with the number of iterations depending on the complexity of the environment. The system successfully converged on the desired focal point even with incomplete information (using 1D electric field data). The CNN-based FSN effectively predicted phase compensation, with MAE losses close to 0. The Adam-based controller efficiently adjusted voltages to achieve focusing. The system's adaptability stemmed from the continuous learning and self-correction capabilities of the SEL algorithm. The use of a fuzzy gradient in the voltage calculation mitigated the impact of uncertainties related to the phase-voltage response curve and the FSN.

The findings demonstrate the effectiveness of the SELAF system in achieving real-time adaptive focusing. The use of SEL addresses the limitations of traditional deep learning approaches by enabling continuous adaptation to changing environments without relying on extensive prior training data. The system's successful operation in complex scenarios highlights its robustness and potential for various applications. The bio-inspired design approach, drawing upon the human eye's accommodation mechanism, offers a new paradigm for developing adaptive optical systems. The results suggest that incorporating biological principles into the design of artificial systems can lead to improved performance and adaptability.

This study successfully demonstrated a bio-inspired neuro-metasurface focusing system using a novel supervised evolving learning algorithm. The system exhibits real-time adaptive focusing capabilities in complex electromagnetic environments, surpassing the limitations of traditional deep learning-based methods. Future research could explore the application of this framework to other areas of electromagnetic wave manipulation, such as 6G communication and advanced imaging systems. The integration of more sophisticated sensors and control algorithms could further improve the system's accuracy and speed.

The current system uses 1D electric field data, which might lead to multiple focal points. While the main lobe energy always dominates, using 2D data could improve accuracy. The focusing efficiency, while reasonable, could be improved through further optimization of the metasurface design and the SEL algorithm. The experimental setup was relatively simple; testing in more complex and realistic environments would provide a more comprehensive evaluation.