Machine Learning Enhanced Post-selection in Quantum Networks

Abstract

Quantum networks are being developed to support secure communication and distributed computation, but their performance is limited by noise in both local operations and transmission. Quantum error correction (QEC) is the standard tool for managing noise, and post-selection adds a frontend stage to the QEC pipeline: instead of decoding every received round, the receiver places a filter in front of the decoder that accepts only trustworthy syndromes for decoding and asks the sender to retry on the rest. In this paper, we examine how machine learning can improve the post-selection process modeled by the filter. We design two such neural network filters, a 3D-CNN and a transformer with FiLM noise conditioning, and evaluate them using the rotated surface code together with an AlphaQubit-style neural network decoder. As a baseline we take the filter built from the decoder's own soft output, which is itself already a strong post-selection signal; against this baseline, our two filters together cut the conditional logical error rate by roughly 2-10× at 10-20% acceptance. Taken together, these results position machine learning enhanced post-selection as a concrete link-layer mechanism for quantum networks, with the practical caveat that the right architecture depends on the operating noise regime.