arXiv:2610.03869
Modular Neural Decoder for Surface Code Memory and Logic on Superconducting Processors
Moritz Lange, Ole Fjeldså, Gustaf Jonasson Johansson, Mats Granath
Abstract
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.
AI Summary
The paper asks whether one neural-decoder architecture can handle surface-code memory, stability measurements, and lattice surgery across superconducting devices and noise settings. A shared graph neural network encodes local $3\times3\times2$ syndrome blocks; convolutional aggregation combines tiles, and a recurrent network predicts an observable flip. Training proceeds through simulated-noise pretraining, calibrated-noise simulation, and hardware-data fine-tuning, with held-out tests. Most models contain roughly 1.6–2.3 million parameters and train on one GPU; configurations are separately fine-tuned rather than decoded zero-shot. On IBM memory experiments, fitted logical error rates per round are 1.26–1.47% at $d=3$, versus 2.60–2.79% for calibrated MWPM, and 2.48% at $d=5$, versus 3.79%. These IBM patches do not exhibit distance-dependent error suppression. On public Google $Z$-memory data, rates are 0.638%, 0.330%, and 0.179% at $d=3,5,7$, respectively, giving $\Lambda\approx1.9$. The $d=7$ result uses a 30-million-parameter variant and remains 38% above AlphaQubit’s error rate. For IBM stability experiments, relative error reductions over MWPM increase from 15% at five rounds to 45% at 100 rounds. One decoder jointly fine-tuned across 17 duration/measurement-noise settings matches or exceeds separately calibrated MWPM. A two-patch $d=3$ lattice-surgery test achieves 58.2% merge/final-parity consistency, versus 56.2% for MWPM and 54.2% uncorrected; it tests one logical correlation, not complete operation fidelity. Stated limitations include increasing recurrent-state bottlenecks with distance, Google tests restricted to the $Z$ basis, hardware-noise-limited IBM experiments, and no latency or throughput benchmarks establishing real-time operation.
Error found: The distance-7 capacity comparison does not isolate recurrent capacity as claimed: the appendix specifies that the wide model also increases spatial aggregation width from 256 to 1024, contradicting the statement that its spatial modules are unchanged.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.