Spectral Shaping of Electro-Optical frequency Combs using Machine Learning Techniques

Electro-Optical frequency combs (EOFC)s are widely applied in areas such as metrology, spectroscopy, light detection and ranging, detection of exoplanets, and optical
communications systems. Among their advantages, EOFCs offer high stability in the frequency spacing, coherence in the phases, and the possibility to replace an array
of lasers with one single optical source. In contrast, EOFCs require strict control over the power distribution per carrier (or line), constituting an essential parameter
for efficient EOFC applications. The EOFC power profile is mainly represented by the flatness which is the difference between the maximum to minimum peak power
in a frequency window. In this work, machine learning (ML) algorithms are demonstrated to improve the characteristics of EOFCs in terms of flatness, carrier-to-noise
ratio (CNR), and the number of carriers. The laser is driven by optimized bias current and RF driving signal composed of multiple harmonics. The bias current,
amplitude, and relative phases of the harmonics in the laser RF driving signal are optimized using inverse system design, reinforcement learning, and gradient-free optimizers such as particle swarm optimization (PSO) and differential evolutionary (DE) algorithms. The methods presented in this work support contributions for developing an optimization benchmark for EOFCs, automation of setups including EOFC combs, and use of harmonic composition to improve flatness in combs. The contributions are demonstrated numerically and experimentally using EOFCs based on GS-lasers, Mach-Zehnder modulators, and silicon ring resonator modulators. The optimization techniques yield to improve performance of state-of-art EOFCs, and the results in this work show how ML techniques support the increase of capabilities applications of EOFCs.