Mitigating Errors Caused by Polarization Drift in Point-to-Point Quantum Communication Using an Artificial Intelligence-Based Pre-Decoder

Polarization-encoded quantum communication over optical fiber is degraded by polarization drift, which rotates the transmitted state before it reaches the receiver and raises the quantum bit error rate (QBER). Existing countermeasures realign the polarization in the optical domain before detection. Convolutional pre-decoders for quantum error correction work instead on measured syndromes, but they are trained for one code and one noise model. This paper poses the question whether a pre-decoder trained on surface-code syndromes can be transferred to drift-induced errors once those errors are expressed in a compatible tensor representation. We formalize the drift channel in the Jones calculus and derive the per-transmission error probability under an arbitrary polarization rotation. The adaptation task is then stated as a tensor-to-tensor classification problem whose objective is the residual QBER. The Fagundes pre-decoder method maps simulated transmission records into synthetic 4×9×9×9 syndrome tensors and fine-tunes every layer of the original pretrained Chamberland pre-decoder Chamberland et al. [2026] on them. Two simulation experiments with separatedatasets and checkpoints were conducted to evaluate the method. Experiment I applied a bounded angular random walk inside the SeQUeNCe simulator at twelve drift intensities. Experiment II applied the Czegledi stochastic Jones-matrix process at six correlation times between 10 s and 0.1 s. In Experiment I the adapted model lowered QBER for drift intensities between 0.05 rad and 0.70 rad, with a peak reduction from 4.87% to 3.80% at 0.40 rad. In Experiment II it lowered QBER from 9.84% to 8.53% on an independent test set of 600 trajectories, which corresponds to a net removal of 74,622 errors. It improved every one of those trajectories, with a bootstrap 95% confidence interval of 1.25 to 1.37 percentage points for the mean reduction. The unadapted original pre-decoder raised QBER in both experiments, so the improvement is associated with the proposed adaptation. These results show that transfer learning converts a surface-code pre-decoder into a drift-error mitigator within a bounded operating range, and they quantify where that range ends.