Intelligent Mode Sorting in Turbulence with Task-Dependent Optical Neural Networks
The practical deployment of high-capacity free-space optical communication is fundamentally limited by atmospheric turbulence, a challenge that conventional mode-sorting techniques have failed to overcome. While engineered optical computers like diffractive networks offer a potential solution, their design complexity remains a significant barrier. Here, we introduce and experimentally validate a methodology for task-dependent hardware design, demonstrating a simple, continuous-wave-driven multimode fibre reservoir that is physically configured to solve this specific, high-impact problem. We first establish a set of physical design principles, showing that recurrent dynamics are optimal for structural data (MNIST), whereas high mode-mixing is superior for textural data (FMNIST). By treating turbulence-induced wavefront distortion as a complex textural feature, we configure a physically optimised reservoir for classifying orbital angular momentum (OAM) modes. In moderate to high-turbulence regimes, our system outperforms an ideal modal decomposition by an average of 20.32~$\pm$~3.00\%, succeeding precisely where the conventional approach fails. This work not only demonstrates a practical, low-complexity solution for turbulence mitigation but also reframes the optical receiver as a physical likelihood processor, thereby offering a path towards significantly reduced digital signal processing burdens by offloading much of the computational load to the physical optical front-end. We establish a framework for co-designing physical hardware with computation, enabling simpler, more robust optical machine learning systems tailored for real-world challenges.