A DeepONet Framework for Enhanced Surrogate Modeling of Reconfigurable Intelligent Surfaces With Limited Data

Reconfigurable intelligent surfaces (RISs) have attracted significant attention due to their abilities to dynamically manipulate electromagnetic wave propagation. In this work, an efficient surrogate modeling framework based on an improved Deep Operator Network (DeepONet) is proposed to predict the far-field radiation patterns of RIS, significantly reducing the reliance on computationally expensive full wave simulations by the online inferences. Specifically, the proposed architecture integrates a convolutional branch network to extract spatial features from RIS configurations and a multilayer perceptron (MLP) based trunk network to encode angular information. Outputs of both networks are fused and passed through a decoder to generate accurate radiation pattern predictions. We verify the proposed framework on the RIS that employs a 20×20 array of 1 bit controllable units, with binary states representing the ON/OFF conditions. It is shown that, compared to several state-of-the-art learning frameworks, the proposed model suggests superior prediction accuracy and training efficiency under limited data scenarios, offering a promising solution for fast and reliable RIS radiation pattern evaluation.