Deep learning prediction of noise-driven nonlinear instabilities in fibre optics.

Abstract

Machine learning is bringing revolutionary approaches into many fields of physics. Among those, photonics enables fast and scalable information processing. Photonics platforms further possess rich nonlinear dynamics that drive fundamental interest but also prove powerful for applications in computation, imaging, frequency conversion, source development and advanced signal processing. However, incoherent processes of nonlinear optics are hardly exploited in practice as the control of noise-driven dynamics remains challenging. Here, we exploit deep learning strategies and demonstrate that coherent optical seeding can effectively shape incoherent spectral broadening. We focus on the intricate interplay between weak coherent pulses and broadband noise, competing during nonlinear fibre propagation within an amplification process known as modulation instability. We demonstrate artificial neural networks' capability to efficiently predict these complex incoherent dynamics, both numerically and experimentally. Our results show that input seed properties can be inferred from the incoherent output signal. Furthermore, our approach enables reliable prediction of output spectral fluctuations, paving the way to tailoring complex photonic signals with specific correlation features.