We introduce a physics-informed graph neural network decoder for the surface code. A discrete Poisson equation on the syndrome graph acts as a hard inductive bias on which a graph encoder learns syndrome-adaptive edge weights. A differentiable solver returns the node potentials and the associated edge current. Our central observation is that the logical-error signal is not carried by any part of that current, but by how the syndrome sits relative to the two code boundaries. We place a sink on each of the two boundaries linked by the logical operator and read the difference of the currents they drain. We prove that this single number is a topological pairing: it weighs each excited detector by a smooth coordinate that runs from +1 on one boundary to -1 on the other, and sums the votes. The coordinate is fixed by both the code topology and the learned metric, generating an exact and free-parameter readout. On the rotated surface code under circuit-level depolarising noise, this one topological scalar matches the best full-field readout. That parity is itself the result: for one logical qubit the logical signal is one-dimensional, so projecting onto it discards nothing. At larger code distance the picture strengthens indicating that isolating the pairing helps more as the field grows larger and sparser.
We introduce a physics-informed graph neural network decoder for the surface code. A discrete Poisson equation on the syndrome graph acts as a hard inductive bias on which a graph encoder learns syndrome-adaptive edge weights. A differentiable solver returns the node potentials and the associated edge current. Our central observation is that the logical-error signal is not carried by any part of that current, but by how the syndrome sits relative to the two code boundaries. We place a sink on each of the two boundaries linked by the logical operator and read the difference of the currents they drain. We prove that this single number is a topological pairing: it weighs each excited detector by a smooth coordinate that runs from +1 on one boundary to -1 on the other, and sums the votes. The coordinate is fixed by both the code topology and the learned metric, generating an exact and free-parameter readout. On the rotated surface code under circuit-level depolarising noise, this one topological scalar matches the best full-field readout. That parity is itself the result: for one logical qubit the logical signal is one-dimensional, so projecting onto it discards nothing. At larger code distance the picture strengthens indicating that isolating the pairing helps more as the field grows larger and sparser.
Quantum error correction is essential to achieve fault-tolerant quantum computation in the presence of noisy qubits. Among the most promising approaches to quantum error correction is the surface code, thanks to a scalable two-dimensional architecture, only nearest-neighbor interactions, and a high error threshold. Decoding the surface code, i.e. finding the most likely error chain given a syndrome measurement outcome is a computationally complex task. Traditional decoders rely on classical algorithms, which, especially for larger systems, can be slow and may not always converge to the optimal solution. This thesis presents a novel approach to decoding the surface code using graph neural networks. By mapping the syndrome measurements to a graph and performing graph classification, we find that the graph neural networks can predict the most likely error configuration with high accuracy. Our results show that the GNN-based decoder outperforms the classic minimum weight perfect matching (MWPM) decoder in terms of accuracy. With a phenomenological noise model with depolarizing noise and perfect syndrome measurements, our networks beat MWPM up to code-size 15 across all relevant error rates. Furthermore, the GNN is capable of surpassing MWPM under circuit-level noise up to code size 7. We also show that training the network on repetition code data from a recent experiment [Google Quantum AI, Nature 614, 676 (2023)] produces per-step error rates comparable to those achieved with a matching decoder specifically adapted to the error rates of the physical qubits. This indicates that graph neural network decoders are capable of learning the underlying error distribution on the qubits. Our findings advance the field of quantum error correction and provide a promising new direction for the development of efficient decoding algorithms.
To utilize quantum error-correcting codes, a decoder must infer the logical sector from the measured syndrome. Beyond a hard logical decision, some decoders provide soft information estimating its reliability. For minimum-weight perfect matching (MWPM), a common confidence measure is the complementary, or logical, gap. Neural-network decoders trained with binary cross entropy naturally output a different score, the logit, whose relation to the matching gap remains unclear. Here, we test whether the logit of a graph neural network (GNN) decoder behaves as a learned proxy for the logical gap. Using a pretrained GNN for the rotated surface code under uniform circuit-level noise [PRR, 7(2):023181, 2025], we compare its logit with the MWPM gap shot by shot on identical syndromes. The native GNN decoder-logit pair yields a lower post-selected logical error rate than the native MWPM decoder-gap pair over the operationally relevant range of acceptance rates. Exchanging the scores between decoders isolates score discrimination from hard-decision accuracy and shows that both contribute to the advantage. At larger distances each score best ranks failures of its own decoder, whereas at small distance, where the bounded matching gap saturates, the learned confidence better discriminates MWPM failures than the gap itself. Finally, GNN confidence is substantially better described by the posterior logistic form implied by an ideal log-likelihood ratio, while the MWPM gap shows systematic departures. Thus, a neural-network decoder trained only on syndromes and logical labels learns a gap-like confidence proxy whose native scale more closely follows the expected posterior calibration form, supporting learned soft output when MWPM gap estimates are unavailable, costly, or insufficiently expressive.
Fault-tolerant quantum computing critically relies on quantum error correction. We study surface-code decoding with diagonally correlated two-qubit Z errors, a structured noise model that breaks assumptions leveraged by MinimumWeight Perfect Matching (MWPM). We develop a graph neural network (GNN) decoder based on learned message passing and benchmark it against a tuned MWPM baseline informed by the same noise statistics. In code-capacity simulations at code distances 3, 5, and 7, the GNN achieves consistently lower logical error rates than MWPM across the tested range of pair-error probabilities. The GNN correction threshold is about 2.5%, surpassing the 2.4% achieved by the MWPM baseline. These results indicate that data-driven message passing can adapt to complex, geometry-dependent noise without hand-crafted correlation models, thereby reducing reliance on manual model specification for quantum error correction.
We present POLYMECHANON, a graph neural network (GNN) decoder for quantum error correction whose only input is the detection error model (DEM) of a quantum code under a given noise model. We represent the DEM as a tripartite graph of detectors, error mechanisms and logical observables, where every input feature is computed by using the quantum code as data rather than design choice. In this way, the same architecture decodes in principle any stabiliser code, under any noise model that can be expressed as a detection error model, once trained on the DEM. We test this approach along four directions. Firstly, on the rotated surface code the decoder outperforms correlated MWPM under both phenomenological and circuit-level noise, with up to 25% fewer logical failures. Secondly, on a family of high-rate qLDPC codes it matches BP+OSD on the smaller codes and surpasses it on the larger ones, with up to 16% fewer logical failures on [[130,4,6]]. Thirdly, its decoding time does not depend on the physical error rate, whereas that of BP+OSD grows with it: on [[130,4,6]] its effective cost per shot is of the order of a few ms on a single GPU against a few tens of milliseconds for BP+OSD on a single CPU core, and its fixed computational graph makes it a candidate for real-time decoding on neutral-atom processors. Finally, a single model trained across codes of different families outperforms uncorrelated MWPM on the graph-like codes seen during training and stays within 10% of BP+OSD on the others, while it tends to fail to generalize to unseen codes, particularly larger ones. Its probabilistic output enables confidence-based post-selection, lowering the logical error rate by more than an order of magnitude on the surface code while keeping more than 90% of the shots. Since only the DEM changes, new codes and noise models can be decoded without redesigning the decoder.
Abstract. Suppressing qubit errors is one of the main challenges in building fault-tolerant quantum computers, requiring quantum error correction to protect fragile quantum information during computation cycles. The surface code is a leading approach because of its high error tolerance and compatibility with realistic hardware, but it requires a decoder that can accurately and quickly use the syndrome measurements to discern whether a logical error has occurred. This study investigates whether a feedforward neural-network can serve as a practical decoder for the surface code under circuit-level noise. Using the Stim software package, this study generates surface code circuits at multiple code distances and physical error rates and compiles them into detector samplers, then trains a PyTorch neural-network to predict whether a logical observable flip occurred from the syndrome data. The neural-network decoder is compared against Minimum-Weight Perfect Matching (MWPM), focusing on metrics including logical error rate and inference speed. The results show that the neural-network decoder performs competitively with MWPM for code distances 3 and 5 and provides inference time advantages at higher physical error rates as its inference time remains relatively constant over increasing physical error rates. However, MWPM outperforms the neural-network at distance 7, suggesting that scalability remains a major limitation for neural-network decoders. These results show that neural-networks are a promising approach for tackling quantum error correction, although future work is still needed to improve these models for real systems at larger code distances.
Neural decoders for quantum error correction have achieved high accuracy on surface code memory experiments, but practical fault-tolerant computation requires decoders that transfer across devices, noise conditions, and logical operations. We introduce a modular graph-recurrent decoder designed for this setting. Built from reusable local syndrome encoders, it is pretrained in simulation and fine-tuned to hardware using a common pipeline, with only lightweight spatial aggregation adapted between experiments. The decoder performs competitively on conventional memory benchmarks, outperforming minimum-weight perfect matching (MWPM) at distances 3 and 5 on an IBM Nighthawk-generation processor and closely approaching AlphaQubit on Google's below-threshold hardware data. We then test the framework in two settings beyond memory decoding on the same IBM device. First, in an X-type stability experiment, we continuously tune the effective measurement error rate using a variable pre-readout rotation and train a single decoder jointly across the resulting noise range. It matches or outperforms MWPM across the sweep without requiring a separately calibrated decoder for each setting. Second, we decode a ZZ-merge lattice-surgery experiment between two distance-3 surface-code patches, improving the decoded correlation between the merge outcome and the final ZZ parity relative to MWPM on real hardware. Together, these experiments demonstrate neural decoding that is portable not only across hardware, but also across qualitatively different surface code observables and logical operations, using models trainable on single-GPU academic resources.