Metasurface modeling method based on feature fusion convolutional neural network

To address the issue of low prediction accuracy resulting from input and output dimension mismatch in deep learning-based metasurface design, a metasurface modeling method based on data fusion convolutional neural networks is proposed. Firstly, the software co-simulation is utilized to model and simulate the structure of the superstrate, obtaining a dataset consisting of one-dimensional structural parameters, two-dimensional images of the superstrate, and corresponding electromagnetic responses. Then, a CNN is constructed to extract the features of the two-dimensional images of the superstrate, which are combined with the one-dimensional structural parameters to form a new feature matrix. This feature matrix is then trained with a fully connected layer to predict the S₂₁ parameters of the superstrate with the forward network and to predict the structural parameters of the superstrate with the inverse network. The trained model has good prediction capabilities, with mean squared errors of 3.148×10⁻⁴ and 2.548×10⁻³ for the forward and inverse network models, respectively. Further optimization of the network model is achieved through genetic algorithms, and transfer learning is utilized to accelerate the training process with new datasets. The results show that the algorithm-optimized network effectively avoids a large number of modeling operations and numerical computations, while maintaining high computational accuracy.