arXiv:2610.01683
A Code-Agnostic Graph Neural Network Decoder from the Detection Error Model
Federico Alberto Astolfi, Guido Pupillo
Abstract
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.
AI Summary
The paper asks whether a neural quantum-error-correction decoder can use a common architecture across stabiliser-code families and noise models by operating directly on a detection error model (DEM). POLYMECHANON represents the DEM as a tripartite graph of detectors, error mechanisms and logical observables. A recurrent message-passing GNN uses syndromes, fault priors and DEM-derived features, including detector coordinates, to predict logical flips. Most benchmarks train separate weights for each code and noise model. Simulated memory-$Z$ experiments use $r=d$ syndrome rounds and compare decoders on identical shots. Against correlated MWPM, the GNN reduces logical failures by approximately $16$–$24\%$ on rotated surface codes with $d=3,5,7,9$ under phenomenological noise, and $10$–$25\%$ for $d=3,5,7$ under circuit-level noise. On phenomenological LaCross codes encoding four logical qubits, performance progresses from parity with BP+OSD at $[[34,4,3]]$ to $16\%$ fewer failures at $[[130,4,6]]$. It remains approximately $4\%$ worse on $[[65,9,4]]$. Confidence-based post-selection reduces surface-code logical errors by factors of $14$–$40$, retaining over $93\%$ of shots for $d\ge5$ at the reported operating point. Inference has fixed, syndrome-independent computation. For $[[130,4,6]]$, effective per-shot cost is $5.1$ ms on an H200 GPU versus $26$–$38$ ms for BP+OSD on one CPU core; this is not a hardware-matched comparison or a streaming demonstration. A shared model trained on fifteen codes approaches classical baselines on training codes, but transfer deteriorates on larger unseen graphs and fails from phenomenological to circuit-level noise. Real-time implementation and broader generalisation remain open.
Central claim may not hold: The claimed applicability to arbitrary DEMs exceeds the documented feature construction, which assumes coordinate and measurement-round conventions that arbitrary DEMs need not provide. Noise-independent forward-pass cost does not imply noise-independent average decoding cost when zero-syndrome shots bypass the network. The fitted low-noise exponents do not establish near-optimal decoding because they leave the logical-error prefactors and gap to maximum-likelihood decoding undetermined. The post-selection explanation incorrectly requires physical error weight at least d for logical failure, although failure can occur at weight ceil(d/2).
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 19 pages, 8 figures, 6 tables