Data-Driven SOA Parameter Discovery and Optimization Using Bayesian Machine Learning With a Parzen Estimator Surrogate

Semiconductor optical amplifiers (SOAs) are building blocks of several active photonic integrated circuits such as lasers and all optical switches. However, the optimization of SOAs is a computationally expensive task due to the high dimensionality of the problem (i.e SOA length, Recombination Parameters, gain coefficient, among others) and the relative long computational time of each simulation run. Furthermore, to accurately simulate optical networks with cascaded SOAs based optical switches one must have access to an accurate model of the device which is not always available. In this work we use a Bayesian Optimization approach based on the single and multi-objective tree-structured Parzen Estimator (TPE) algorithm to find parameters for two wideband models of SOAs operating in different parts of the optical spectrum, the first one in the O-band, and the second one covering the S, C and L-bands. With less than 100 function evaluations and on a limited amount of training (measured) data we are able to obtain general models of both SOAs with a worst average error of 1.12 dB for the gain and -1.86 dB for the optical signal-to-noise ratio (OSNR) in the O-band SOA and a worst average of 0.64 dB for the gain and 0.81 dB for the OSNR in the C-band SOA. We also found that the presented approach outperform common used evolutionary algorithms and Gaussian Processes based Bayesian optimization with regard to the number of required function evaluations, with the TPE obtaining a mean squared error (MSE) of -27 after just 13 trials and the second best, an evolutionary algorithm, obtaining a minimum MSE of 33 after 40 trials.