Physics-Informed Machine Learning Optimization Design of Reconfigurable Intelligent Surfaces for IoT Applications

Reconfigurable intelligent surfaces (RISs) is a two-dimensional metamaterial integrated with numerous reflective RIS elements, which manipulate electromagnetic (EM) waves digitally, offering great potential for internet of things (IoT) applications. However, conventional designs rely heavily on extensive full-wave EM simulations, which are extremely time-consuming. To address this challenge, we propose a machine-learning-assisted approach for efficient RIS design. First, an accurate and fast model for predicting the reflection coefficient of a RIS element is developed by integrating a multi-layer perceptron neural network with a dual-port network. This integration significantly reduces reliance on tedious EM simulations during training, enhancing computational efficiency. Once established, the model serves as a computationally efficient surrogate for resource-intensive EM simulations, enabling streamlined RIS element parameter optimization. Subsequently, an efficient framework for optimizing RIS element design parameters is proposed, advancing the design process. To validate the method, three RIS elements with distinct center frequencies were designed for IoT Applications. Among these, one RIS element was selected to construct a functional RIS. The performance of the RIS was evaluated through fabrication and experimental measurements. Notably, the experimental results demonstrate excellent agreement with EM simulations, confirming the effectiveness of the proposed method.