Simulation and Analysis of Optical Neural Network Accelerators Based on Multimode Interferometers and Phase Shifters

Optical unitary converters, which are implemented by multiport directional couplers or multimode interferometers, offer compactness and lower-power operation when performing unitary matrix operations. Multimode interferometers are particularly advantageous due to their good tolerance to fabrication errors and insensitivity to wavelength and polarization. Moreover, cascading optical unitary converters with Mach-Zehnder Interferometer arrays can achieve arbitrary matrices, which are frequently used in neural network computation. In this paper, we simulate and analyze an optical neural network accelerator based on multimode interferometers and phase shifters to perform arbitrary matrices. We compare its performance under different phase deviation conditions and evaluate its feasibility in neural network computations through convolutional operations. The simulation results demonstrate that the mean square error of the proposed accelerator is approximately -20dB and it exhibits excellent performance in realizing convolutional operations.