Studying the statistical properties of chaotic semiconductor lasers by machine learning

The statistical properties associated with chaotic semiconductor lasers have been widely investigated for different applications. Traditional methods usually rely on characterizing the statistics from direct measurement of a temporal emission waveform, which is usually recorded in an electrical domain after optical-to-electrical conversion. In this work, we propose a machine learning-based methodology to study the statistical properties by measuring merely the optical spectrum of a chaotic emission output. Numerical simulations are first conducted to verify the feasibility of the proposed method based on a chaotic optically injected semiconductor laser. By utilizing a feed-forward neural network, our method trains on optical spectrum data to predict the local maximum peak intensity of a chaotic emission waveform, thereby enabling successful characterization of the associated statistics. The flexibility of the proposed method is validated by varying the operating parameters of the machine learning model and the injection parameters for laser chaos generation. The impacts of practical spectral resolution bandwidth, laser inherent noise, measurement noise, and laser parameter mismatch are then investigated in detail. This work provides a new perspective for studying the statistical properties of chaotic semiconductor lasers.