Implementation and Reinforcement Learning-based stabilization of a digital twin of a loop-based photonic quantum computer

Loop-based photonic architectures offer a scalable approach to quantum computation by reusing a single interferometer for multiple time-bin modes. However, a major limitation of such systems is dephasing, which consists of the emergence of parasitic phase shifts within the interferometer, which in turn may compromise the computation.

To address such issues, we developed a digital twin of a loop-based photonic quantum computer, able to classically emulate Gaussian Boson Sampling processes and benchmark state-of-the-art photonic simulators. Such a classical emulator is built upon a backend in which the unitary interferometric matrix is decomposed through the Reck decomposition. Optical losses can be incorporated through loss factors that model photon transmission between consecutive optical elements.

Relying on the backend developed for the digital twin, we implemented a Gaussian simulated circuit able to reproduce the interferometric dynamics affected by parasitic phase shifts. Setting this environment, a Reinforcement Learning agent based on the Deep Q-Network algorithm was trained to detect and correct such a dephasing in real time. The trained agent successfully minimizes phase errors across multiple episodes, achieving complete success.

By enabling autonomous phase stabilization and realistic loss modeling, our framework provides a versatile testbed for optimizing control strategies before implementation on physical hardware.