Automatic Synthesis of Broadband Silicon Photonic Devices via Bayesian Optimization

In integrated silicon photonics, broadband devices are of high interest. In this paper, we propose a fully automatic design methodology to synthesize broadband silicon photonic devices via Bayesian optimization. Bayesian optimization is a surrogate-model based black-box global optimization technique, made up of two steps in combination: (i) building a surrogate model, and (ii) optimizing a user-defined acquisition function. Key is that the surrogate model serves as a good substitution for the time-consuming physical simulation, while being much cheaper to invoke. Our simulation results verify the efficacy of Bayesian optimization for designing various broadband silicon photonic passive devices over other baseline methods (e.g., L-BFGS-B and differential evolution) under our hyper-parameter settings and using our implementation. Specifically, we show that with around 100 FDTD simulations, far fewer than previously reported in the literature, Bayesian optimization can already attain devices with superior performance. We further demonstrate the trade-off relation between device robustness under variation and device performance.