Inverse Design of Broadband Polarization Conversion Metasurface Based on Residual-Enhanced Neural Network

The rapid development of artificial intelligence provides a customized solution for the free manipulation of electromagnetic waves using metasurfaces. This paper proposes a deep fully connected neural network model incorporating the concept of Residual Networks (ResNet) to inversely predict the structural parameters of a broadband polarization conversion metasurface from reflection coefficients. Firstly, a three-layer metasurface unit structure is designed, and its eight control parameters are determined. Building on this, by combining the concept of precise parameter tuning for different metasurface structures with the efficient mapping capability of deep learning inverse design, an end-to-end mapping model from electromagnetic response to structural parameters is constructed. A residual connection mechanism is innovatively introduced, effectively addressing the vanishing gradient problem during deep network training. The paper elaborates on the residual-fused network architecture design, training strategy, and data preprocessing pipeline. Algorithm evaluation shows that the coefficient of determination (R²) for the predicted results of all eight structural parameters exceeds 0.9. Metasurfaces designed based on the predicted parameters maintain a polarization conversion ratio above 90% across the entire frequency band of 8.8-24.4 GHz. Analysis demonstrates that this study provides an efficient and feasible method for metasurface inverse design. Reasons for performance differences in predicting various parameters are discussed based on data characteristics. The method can be further extended to the design of metasurfaces with more functionalities.