Uncertainty-Aware Neural Decoding for Surface-Code Quantum Error Correction
By Andrew, Lucas
Introduction: Quantum computers must detect and correct physical errors without directly measuring the encoded quantum state. Surface codes address this problem by repeatedly measuring local stabilizers on a two-dimensional qubit lattice. Changes in those measurements, called detection events, reveal where an error may have occurred, but they do not identify a unique physical-error pattern. A decoder must instead choose any correction that returns the state to the correct logical equivalence class. Minimum-weight perfect matching (MWPM) is a strong standard decoder, while recent neural decoders show that learned models can exploit spatial, temporal, and device-specific error correlations.
This project will test one narrow question: does preserving a neural model's uncertainty about physical errors improve surface-code decoding? A spatiotemporal neural network will map repeated syndrome measurements to per-qubit probabilities for the $X$ and $Z$ components of Pauli errors. A second, syndrome-consistent correction stage will receive those probabilities rather than a binary error map. We will compare this soft-information pipeline with MWPM, a direct neural decoder, and an otherwise matched model whose probabilities are thresholded before correction. The primary outcome will be logical error rate, not exact physical-error classification accuracy.
Intellectual Merit: The project will isolate the value of uncertainty through controlled ablation rather than evaluating only whether a neural decoder can correct errors. It will distinguish improvements caused by retaining probability information from improvements caused by model size, a learned syndrome representation, or a strong conventional decoding stage. The study will also compare physical-error inference metrics with the logically meaningful outcome, testing whether better per-qubit predictions actually produce fewer logical failures.
Broader Impact: Reliable decoding is necessary for useful fault-tolerant quantum computation, including future quantum-machine-learning workloads. A reproducible benchmark showing when soft neural information helps, has no effect, or hurts would help researchers avoid judging decoders by misleading physical-error accuracy alone. The data pipeline, ablation protocol, and logical-error analysis could also be reused by future QLab projects studying other codes, noise models, or quantum devices.