We present an open-source deep artificial neural network (ANN) model for the accelerated design of polarization-insensitive subwavelength grating (SWG) couplers on the silicon-on-insulator platform. Our model can optimize SWG-based grating couplers for a single fundamental-order polarization, or both, by splitting them counter-directionally at the grating level. Alternating SWG sections are adopted to reduce the reflections (loss) of standard, single-etch devices-further accelerating the design time by eliminating the need to process a second etch. The model of this device is trained by a dense uniform dataset of finite-difference time-domain (FDTD) optical simulations. Our approach requires the FDTD simulations to be made up front, where the resulting ANN model is made openly available for the rapid, software-free design of future standard photonic devices, which may require slightly different design parameters (e.g., fiber angle, center wavelength, polarization) for their specific application. By transforming the nonlinear input-output relationship of the device into a matrix of learned weights, a set of simple linear algebraic and nonlinear activation calculations can be made to predict the device outputs 1,830 times faster than numerical simulations, within 93.2% accuracy of the simulations.

artificial neural networks, grating couplers, machine learning, polarization insensitivity, Silicon photonics, subwavelength devices
IEEE Journal of Selected Topics in Quantum Electronics
Department of Electronics

Gostimirovic, D. (Dusan), & Ye, W.N. (2018). An Open-Source Artificial Neural Network Model for Polarization-Insensitive Silicon-on-Insulator Subwavelength Grating Couplers. IEEE Journal of Selected Topics in Quantum Electronics. doi:10.1109/JSTQE.2018.2885486