Tandem Neural Networks for the Inverse Programming of Linear Photonic Processors
José Roberto Rausell‐Campo, Daniel Pérez, Bhavin J. Shastri, Daniele Melati · 2023
Linear programmable photonic integrated processors have emerged as an alternative hardware platform for quantum, deep learning, and microwave photonic systems. Calibration and control of the photonic processor using deep learning techniques has proven to be challenging due to the one-to-many problem. This means that a given functionality of the processor can be achieved using different settings of the controllers. In this paper, we demonstrate how tandem neural networks can overcome this limitation in meshes of Mach-Zehnder interferometers by employing forward and inverse networks. This approach is independent of the mesh architecture and introduces a novel method for controlling photonic linear processors using deep neural networks. We provide an experimental demonstration using a 3x3 linear processor, achieving a control resolution higher than 7 bits.