Integrating CBAM-CNN architectures with K-means clustering algorithms for high-efficiency and accurate metasurface optical properties’ prediction

To address inefficient feature extraction and imbalanced data in deep learning for metasurface design, we propose a physics-informed, dual-strategy framework. It integrates a convolutional block attention module into a convolutional neural network model, accelerating feature extraction tenfold, and a K-means clustering algorithm to optimize data distribution. This approach boosts the ratio of predictions with a final loss below 1 from 18% to 91%. We demonstrated its utility by designing a high-performance beam deflector, showcasing a powerful strategy for developing compact optical devices in augmented reality and integrated photonics.