Abstract
Photonic quantum information processing has emerged as a leading platform for demonstrating quantum advantage. Among these, Gaussian Boson Sampling (GBS) is a powerful model of quantum computation. However, practical GBS devices are intrinsically affected by fabrication imperfections, phase noise, and temporal drift, which distort output photon statistics and ultimately limit scalability and long‐term stability. In this work, a noisy GBS device is numerically investigated using a 44 reconfigurable interferometer based on the Clements architecture. The interferometer is modeled using a multimode interference using 3D FDTD, yielding an extinction ratio of 47 dB and an insertion loss of 0.6 dB at 1550 nm. Noise and phase drift are incorporated through a Hamiltonian‐based noise model. To quantify the impact of noise, graph decoding techniques were employed to characterize structural distortions in the sampled output distributions. Further, the fundamental limitations of static calibration in the presence of time‐dependent drift were investigated. This work overcomes these limitations with an adaptive reinforcement‐learning controller that can dynamically compensate for Hamiltonian drift and noise. Which maintains stable output statistics close to the target GBS distribution even under non‐stationary perturbations. Finally, the potential applications of GBS, with the advantage of adaptive stabilization of GBS devices, were discussed.