Chirp-based direct phase modulation of VCSELs managed by Neural Networks

Abstract

VCSEL's capacity of direct modulation and its low cost makes this device a feasible cost-effective transmitter for ultra-dense wavelength division multiplexing (uDWDM) metro-access networks using coherent detection. However, performing direct-phase modulation in semiconductors can be complex due to its nonlinear characteristics. This research presents Neural Network (NN) training techniques for Time-Series analysis in order to describe the correlation between the input current given to the device and its output optical phase, using a 1550nm RayCan SM-VCSEL. Main goal is training a NN capable of predicting an ideal optical power signal for a specific phase result achievable by inverse training, that is: optical phase is the neural network input while the optical power is the desired target. The experiment is done in three stages: (i) VCSEL's characterization, (ii) NN training to predict input current knowing optical power, and (iii) NN training to predict optical power from a known optical phase.