Optical Computing with OAM Modes in Diffractive Neural Networks

Diffractive neural networks (DNNs) are a physical neural network architecture which employ phase-modulating metasurfaces to perform nonlinear computation. Much like how modern graphics processing unit (GPU) cores execute many simple instructions in parallel to optimize machine learning algorithms, DNNs efficiently perform neural network operations by locally phase-modulating discretized wavefronts in successive layers. Structural nonlinearities, which are efficient optoelectronic devices, have been proposed as a method of introducing nonlinearities at each metasurface layer thanks to developments in reconfigurable metasurfaces. As a result, DNNs now show promise and can be simulated to further probe their behavior. We prepare a set of training data with states in the arbitrarily large OAM (Orthogonal Angular Momentum) Mode basis, then determine DNN hyperparameters on that training set through simulation. In simulation, we seek to assess the feasibility of n-dimensional optical gate operations within the OAM mode basis, which could form an architecture for fully-optical computing.