We establish a systematic framework of photonic device discovery using a physics-based deep learning approach. The computationally expensive physics simulations are removed from the critical loop to generate data and perform one-time training of the deep learning models. Consequently, the trained deep learning models achieve massive speed up on the iterative design process. Our approach reduces the computational time from days to minutes. Using a silicon power divider as an example, we demonstrate discovery of a spectrum of devices that simultaneously satisfy compact footprints, ultralow losses, ultrawide bandwidth, and exceptional robustness against fabrication randomness.
|