The Limits of Supercontinuum Generation for Photonic Neuromorphic Computing

There is currently intense effort developing photonic hardware that can perform tasks of artificial intelligence, with impressive results in photonic neural networks, reservoir computers, and extreme learning machines (ELMs) [1]. Several recent studies have specifically shown that the complex dynamics of fibre supercontinuum (SC) generation can emulate the nonlinear transformation of the hidden layer of an ELM [2], [3], but these systems have been optimized experimentally without detailed simulations or digital twin frameworks. As a result, the potential limits of supercontinuum computing remain unknown. Here we address this question using an end-to-end simulation of a supercontinuum ELM that allows us to systematically study how such systems can be optimized.