Deep learning accelerated discovery of photonic power dividers

Abstract

This article applies deep learning-accelerated inverse design algorithms and discovers a spectrum of photonic power dividers with exceptional performance metrics despite the simplicity in the design geometry. The deep learning models exhibit high precisions on the order of 10−6 to 10−8 for both TE and TM polarizations of light. These models enable ultrafast search for an empirically describable subspace that simultaneously satisfy compact footprints, ultralow losses, ultrawide bandwidth, and exceptional robustness against fabrication randomness. We demonstrate a spectrum of devices for silicon photonics with programmable power splitting ratios, excess losses as small as 0.14 dB, to the best of our knowledge, the smallest footprints on the scale of sub-λ 2, and low loss bandwidths covering the whole telecommunication spectrum of O, S, E, C, L and U-bands. The robustness of the devices is statistically checked against the fabrication randomness and are numerically verified using the full three-dimensional finite difference time domain calculation.