Efficient RIS Inverse Design Enabled by Microwave Network Decoupling and Active Learning

Abstract

Reconfigurable intelligent surfaces (RISs) are promising for dynamic wireless networks, yet their inverse design is still limited by costly full-wave datasets and inefficient optimization in high-dimensional topology spaces. To address these challenges, this article proposes a physics-informed and data-efficient inverse design framework. First, a surrogate model combining microwave-network cascading theory with a residual convolutional neural network (CNN) forward predictor is developed. The network predicts the S-parameters of passive RIS topologies, which are then cascaded with diode equivalent impedances to obtain reflection coefficients, enabling rapid prediction for different diode models or bias states without retraining. To reduce data acquisition cost, a query-by-committee (QBC) active learning strategy is introduced to iteratively select high-uncertainty samples for full-wave simulation, reducing the required training data by about 60% while maintaining comparable accuracy. Based on this surrogate, a hierarchical topology optimization method combines nonuniform rational B-spline (NURBS)-based global skeleton search with pixel-level simulated-annealing refinement, preserving structural connectivity while retaining local design freedom. The framework is validated through a 5-GHz 1-bit RIS and a 5.8-GHz 2-bit near-field focusing RIS, demonstrating its efficiency and reliability for rapid reconfigurable RIS design.