# Modeling and Control of a Reconfigurable Photonic Circuit using Deep Learning

## TL;DR

This paper presents a deep learning-based approach to model and control reconfigurable photonic circuits, overcoming challenges of indirect measurement and uncertainties, with applications in classical and quantum control tasks.

## Contribution

It introduces a graybox neural network model combining Hamiltonian estimation and quantum mechanics rules for controlling optical quantum devices.

## Key findings

- Accurately models the photonic circuit with low mean square error.
- Successfully controls classical output power distribution.
- Achieves target quantum gates using neural network-derived control voltages.

## Abstract

The complexity of experimental quantum information processing devices is increasing rapidly, requiring new approaches to control them. In this paper, we address the problems of practically modeling and controlling an integrated optical waveguide array chip, a technology expected to have many applications in telecommunications and optical quantum information processing. This photonic circuit can be electrically reconfigured, but only the output optical signal can be monitored. As a result, the conventional control methods cannot be naively applied. Characterizing such a chip is challenging for three reasons. First, there are uncertainties associated with the Hamiltonian describing the chip. Second, we expect distortions of the control voltages caused by the chip's electrical response, which cannot be directly observed. Finally, there are imperfections in the measurements caused by losses from coupling the chip externally to optical fibers. We developed a deep neural network approach to solve these problems. The architecture is designed specifically to overcome the aforementioned challenges using a Gated Recurrent Unit (GRU)-based network as the central component. The Hamiltonian is estimated as a blackbox, while the rules of quantum mechanics such as state evolution is embedded in the structure as a whitebox. The resulting overall graybox model of the chip shows good performance both quantitatively in terms of the mean square error and qualitatively in terms of the predicted waveforms. We use this neural network to solve a classical and a quantum control problem. In the classical application we find a control sequence to approximately realize a time-dependent output power distribution. For the quantum application we obtain the control voltages to realize a target set of quantum gates. The proposed method is generic and can be applied to other systems that can only be probed indirectly.