Shape-Aware Tandem Neural Networks for Data-Efficient Inverse Design of Plasmonic Bowtie Nanoantennas
Abstract
Developing reliable inverse design networks for complex nanophotonic structures remains challenging due to limited training data, inherent spectral bias and “one-to-many degeneracy” problem. In this work, we propose a data-efficient tandem neural network (TNN) framework that integrates Fourier-feature encoding (FFE) with a custom shape-aware (SA) loss function. We systematically evaluate three deep learning (DL) backbone architectures, including multilayer perceptron (MLP), ResNet, and Transformer, under four different optimization schemes: mean squared error (MSE) loss, SA loss, FFE + MSE loss and FFE + SA loss. Our comprehensive analysis shows that MSE-based optimization can deceptively achieve low spectral reconstruction errors while still producing large errors in geometric parameter prediction. In contrast, the proposed FFE + SA loss scheme substantially improves the accuracy of geometric parameter prediction. Furthermore, we demonstrate that the MLP-based architecture optimized with FFE + SA loss scheme achieves the lowest overall average geometric parameter prediction errors among all tested configurations. This work highlights the inherent challenge in inverse design of complex nanophotonic devices and provides a reliable and robust network framework for mitigating degeneracy in nanophotonic inverse design.