Deep learning enabled inverse design of ultracompact silicon metadevices for mode conversion

Deep learning has shown great potential for enabling the rapid inverse-design of high-performance photonic devices. In this work, we develop a forward prediction and inverse design platform for silicon metadevices for mode conversion using deep neural networks (DNNs). The performances of twenty silicon-based mode converters are predicted by using this platform, with a prediction time of just 5 ms and a root mean square error (RMSE) of less than 0.018. Five silicon-based mode converters were inverse-designed using DNN, achieving an ultracompact dimension of only 4.08 µm × 1.68 µm and a design time of only 10 ms. Experimental results verified the performance of the inversely designed devices and compared them with their forward predictions. The measured insertion losses deviated from the desired and predicted values by less than 0.887 and 0.878 dB, respectively. The DNN-based forward prediction and inverse design approach enables the rapid development of highly integrated and high-performance silicon-based mode converters, providing key components for mode-division multiplexing systems.