Inverse design of 3D integrated high-efficiency grating couplers using deep learning

Abstract In recent years, the use of deep learning to replace traditional numerical methods for electromagnetic propagation has shown tremendous potential in the rapid design of photonic devices. However, most research on deep learning has focused on single-layer grating couplers, and the accuracy of multi-layer grating couplers has not yet reached a high level. This paper proposes and demonstrates a novel deep learning network-assisted strategy for inverse design. The network model is based on a multi-layer perceptron (MLP) and incorporates convolutional neural networks (CNNs) and transformers. Through the stacking of multiple layers, it achieves a high-precision design for both multi-layer and single-layer raster couplers with various functionalities. The deep learning network exhibits exceptionally high predictive accuracy, with an average absolute error across the full wavelength range of 1300–1700 nm being only 0.17%, and an even lower predictive absolute error below 0.09% at the specific wavelength of 1550 nm. By combining the deep learning network with the genetic algorithm, we can efficiently design grating couplers that perform different functions. Simulation results indicate that the designed single-wavelength grating couplers achieve coupling efficiencies exceeding 80% at central wavelengths of 1550 nm and 1310 nm. The performance of designed dual-wavelength and broadband grating couplers also reaches high industry standards. Furthermore, the network structure and inverse design method are highly scalable and can be applied not only to multi-layer grating couplers but also directly to the prediction and design of single-layer grating couplers, providing a new perspective for the innovative development of photonic devices.