We propose and demonstrate the first end-to-end artificial neural network (ANN) modeler for the automated design of photonic systems and devices. This approach gathers an initial range-restricted batch of numerically solved electromagnetic data and maps the nonlinear input-output relationship into a linear model of learned weights. This model is used to predict the output of different device variations for orders-of-magnitude faster optimization or system-level simulations. Our implementation uses the MATLAB numerical computing environment with the finite-difference time-domain electromagnetic solver from Lumerical to acquire the device data, create and train the ANN model, and optimize for a desired device output. In this demonstration, we create a model for a silicon grating coupler, which computes 56,000X faster than the numerical simulation, with an accuracy greater than 97% of the numerical results. Using a parametric sweep or an inverted ANN, the device parameters can be immediately found for a desired output.

15th IEEE International Conference on Group IV Photonics, GFP 2018
Department of Electronics

Gostimirovic, D. (Dusan), & Ye, W.N. (2018). Automating Photonic Design with Machine Learning. In IEEE International Conference on Group IV Photonics GFP (pp. 71–72). doi:10.1109/GROUP4.2018.8478722