Design Features of Optical Diffraction Neural Networks

Recently, significant attention has been focused on finding and implementing approaches that would increase the efficiency of existing computational methods or create fundamentally new ones. One promising direction is the transition from digital to analog computing schemes, which allow for the design of high-performance specialized architectures based on known physical principles. In particular, a physical system in which a structure analogous to an artificial neural network can be implemented is a diffractive neural network. However, transferring computations to an analog platform entails the necessity of precise selection of a mathematical model that adequately accounts for the features of the physical implementation. In this work, the correctness of numerical modeling of a Fourier-diffractive neural network is experimentally tested, and the influence of the system configuration on the accuracy of the final computational result is numerically studied.