Theoretical and simulated analysis of a 6 × 6 MZI-based optical neural network

Mach-Zehnder Interferometer (MZI) topological cascade architectures, which allow linear transformations between multiple channels simultaneously, are gradually becoming a powerful tool for photonics and have a place in optical neural networks. This paper focuses on the theoretical analysis and software simulation of two unitary matrix architectures for a 6×6 MZI-based optical processor: the triangular architecture and the rectangular architecture. Both unitary arrays are composed of fifteen reconfigurable MZI units, each of which is comprised of two adjustable phase shifters and two 3-dB directional couplers. For a given neural network training application, the value of each phase shifter can be calculated from the matrix factorization process. The theoretical derivation is verified through simulations using the advanced software Max-Optics. When compared with Lumerical INTERCONNECT, the maximum error is less than 0.00044%, and the simulation time is reduced by 3 times.