Neural network reconstruction of the reduced transfer matrix for tunable linear-optical chips

In this paper, a method for reconstruction of the reduced transfer matrix for tunable linear-optical chips is proposed without taking into account the complex phases of the matrix elements. In this method there is no need in knowledge about both the architecture of the devices and their error models. The numerical simulation of the proposed method is performed on the example of tunable linear-optical chips of common universal architectures based on the mixing layers and Clements with dimensions of 4 × 4 and 6 × 6. The method was also experimentally tested on the example of a tunable 4 × 4 linear-optical chip with an architecture based on mixing layers. The average fidelity value of the reconstruction of the reduced transfer matrix of the optical chip in the experiment was 91%.