Computational design of efficient grating couplers using artificial intelligence

Abstract

We examined the applicability of neural networks (NNs) and the covariance matrix adaptation evolution strategy (CMA-ES) for designing high-efficiency grating couplers on a Si-on-insulator (SOI) photonics platform operating at a wavelength of 1550 nm. We trained a NN to predict the wavelength dependences of coupling efficiency instead of using the finite-difference time-domain method, which enabled us to calculate analytically the gradient of coupling efficiency with respect to the design parameters. As a result, we found that a gradient-based method can be applied to find a more optimal design from an apodized grating coupler. In contrast, the CMA-ES, which is one of the most efficient evolutionary strategies, enabled us to explore a large design-parameter space efficiently without any constraint in an initial design of a grating coupler. As a result, we successfully obtained the optimal design from a uniform grating coupler. The optimized grating coupler exhibited a coupling efficiency of 74.7% when the thickness of the Si layer is 220 nm. The CMA-ES also enabled us to optimize the thickness of the Si gratings simultaneously. Through the optimization, we achieved a coupling efficiency of 92.1% when the Si thickness is 293 nm.