The impact of nonlinear polarisation coupling in optical fibre-based extreme learning machines.

Optical fibre-based extreme learning machines (ELMs) are a powerful physical platform for unconventional computing in photonics. Recent experiments have demonstrated ELM operation using both anomalous-dispersion supercontinuum generation and normal-dispersion wave-breaking. Building on our previously reported generalized nonlinear Schrödinger equation (GNLSE) based numerical pipeline, we present the first numerical study of polarization-dependent effects in fibre ELMs using coupled vectorial GNLSEs with Raman, self-steepening, and higher-order dispersion. We investigate MNIST classification accuracy versus input polarisation and readout strategies in both dispersion regimes. Our results reveal how polarisation coupling influences ELM performance, providing insight for designing polarisation-sensitive nonlinear fibre computing platforms.