Abstract Fault-tolerant quantum computers must continually decode noisy parity-check data, known as the syndrome, to determine whether an encoded logical qubit has been corrupted. The standard decoder, minimum-weight perfect matching (MWPM), assumes independent, graph-like errors and is sub-optimal under realistic circuit-level noise. We present a compact residual 3D convolutional neural network (0.5–0.6 million parameters) that reads the space-time syndrome of a rotated surface code as a volumetric image. Trained on Stim simulations at code distances d = 3 and d = 5 and evaluated on 3.6 million held-out shots, a three-seed ensemble outperforms correlated MWPM, the strongest baseline we test, at all 18 noise settings, with 6–46% fewer logical errors (paired McNemar p-value below 0.03 in every setting). On real Google Sycamore data (d = 3, five rounds), the network is pre-trained on the device's fitted error model and fine-tuned on half of the shots. It achieves a logical error rate of 13.03%, compared with 13.57% for correlated MWPM on the same held-out shots, and outperforms it on all eight cross-validation folds, although it does not match belief-matching or tensor-network decoders. Its outputs are well calibrated (ROC AUC = 0.998): discarding the 10% least-confident shots reduces errors tenfold. An ablation study identifies space-time locality as the essential inductive bias. All models train in 49 minutes on two freely available GPUs, and the data and code are publicly available.
Abstract Fault-tolerant quantum computers must continually decode noisy parity-check data, known as the syndrome, to determine whether an encoded logical qubit has been corrupted. The standard decoder, minimum-weight perfect matching (MWPM), assumes independent, graph-like errors and is sub-optimal under realistic circuit-level noise. We present a compact residual 3D convolutional neural network (0.5–0.6 million parameters) that reads the space-time syndrome of a rotated surface code as a volumetric image. Trained on Stim simulations at code distances d = 3 and d = 5 and evaluated on 3.6 million held-out shots, a three-seed ensemble outperforms correlated MWPM, the strongest baseline we test, at all 18 noise settings, with 6–46% fewer logical errors (paired McNemar p-value below 0.03 in every setting). On real Google Sycamore data (d = 3, five rounds), the network is pre-trained on the device's fitted error model and fine-tuned on half of the shots. It achieves a logical error rate of 13.03%, compared with 13.57% for correlated MWPM on the same held-out shots, and outperforms it on all eight cross-validation folds, although it does not match belief-matching or tensor-network decoders. Its outputs are well calibrated (ROC AUC = 0.998): discarding the 10% least-confident shots reduces errors tenfold. An ablation study identifies space-time locality as the essential inductive bias. All models train in 49 minutes on two freely available GPUs, and the data and code are publicly available.
Learned decoders are often proposed as a faster, more adaptable replacement for minimum-weight perfect matching (MWPM) in quantum error correction, yet published comparisons mix datasets, noise levels and metrics, and many omit a matching baseline altogether. This paper contributes a controlled benchmark in which a multilayer perceptron (MLP), a two-layer LSTM and a Transformer encoder are trained and tested on identical Stim-generated repetition-code data (distances 3 to 11, circuit-level depolarizing noise, five measurement rounds) under one training protocol, and are scored against MWPM and a trivial majority-class predictor over five independent seeds. We report logical error rate, balanced accuracy, F1 for the logical-error class and ROC-AUC rather than raw accuracy, because roughly 12% of shots at a physical error rate of 0.05 (and under 1% at 0.003) carry a logical flip, so a decoder that never predicts an error already scores 88% (99.2% at the lower rate). With 3,200 training shots, none of the neural decoders matches MWPM at any distance. MWPM drives the logical error rate from 4.3% at distance 3 to nearly zero at distance 11, whereas the neural decoders stay between 5% and 12% and the MLP falls back to the majority prediction once the syndrome has 60 bits. Presenting the temporal structure of the syndrome to the network, one token per measurement round, gives the largest and most consistent gain, and a Transformer fed scalar detector tokens without positional information collapses. We also examine the effect of training-set size, a shift in test-time noise, and decoding latency, and we discuss which conclusions the data do and do not support.
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.
Fault-tolerant quantum computation critically depends on the ability to detect and correct physical qubit errors faster than they accumulate. The surface code (SC) has emerged as the most experimentally promising quantum error correction (QEC) framework due to its high threshold and local stabilizer structure, yet its practical performance is bottlenecked by decoder efficiency. Traditional decoding approaches such as minimum-weight perfect matching (MWPM) achieve near-optimal accuracy under independent noise but struggle with correlated errors, circuit-level noise, and real-time latency constraints. This paper investigates how neural network (NN)-based decoders reduce logical error rates in SC quantum systems by learning complex syndrome-to-correction mappings from data. We propose a hybrid convolutional-recurrent neural architecture that processes stabilizer measurement syndromes across both spatial and temporal dimensions, enabling the decoder to capture correlated error patterns inherent in realistic hardware. Systematic simulations under depolarizing, biased, and circuit-level noise models demonstrate that the proposed neural decoder achieves a logical error rate reduction of up to 41.3% compared to MWPM at a physical error rate of p = 0.008, while maintaining sub-millisecond decoding latency compatible with superconducting qubit cycle times. These results establish that data-driven neural decoders represent a scalable and hardware-compatible pathway toward achieving the logical error rates required for practical fault-tolerant quantum computing.
Abstract One of the critical requirements for fault-tolerant quantum computation is fast and reliable decoding,Minimum-weight perfect matching (MWPM) is a reliable surface-code decoder,but it will face long decoding time when decoding instances grow. Neural-network decoders was used to solve this problem, but large models often require costly retraining when the code distance or noise model changes.In this work, we propose a two-stage decoder that combines pretrained graph neural network (GNN) residual pre-decoding with global MWPM cleanup.The GNN receives a detector graph generated by the established decoding backend and assigns confidence scores to local correction candidates.Two staggered layers of parallel local windows process confidence corrections.The residual syndrome is passed to the NVIDIA-accelerated MWPM backend.On matched batches of 32,768 samples, the complete method reduced the logical failure rate by 6.25% to 18.01% relative to Pure MWPM for code distances d=3,5,7 and 9 at physical error rates p = 0.007 to 0.013. At d=3,it decoded 3.48x – 4.38x faster over p=0.003 to 0.021. We also trained separate GNN residual pre-decoders for four circuit-level noise models over 160 code-distance/error-rate configurations and obtained 2.25x to 5.41x speedups in every case.The logical-failure benefit depended on the noise model,but the pace is all improving. Our work can accelerate surface-code decoding pace and extend to others noise models.
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.
Real-time decoding is a critical bottleneck for large-scale fault-tolerant quantum computing. AI-based neural pre-decoders locally correct most physical errors before passing residual syndromes to a global decoder, enabling sub-microsecond latencies. However, existing architectures carry significant overhead from dense 3D convolutions. We present QuantiSpect, a lightweight 3D convolutional neural network (CNN) pre-decoder for the rotated surface code, built on the decoding pipeline of Chamberland et al. The key idea is to replace the dense 3D convolutions with three parallel branches in each residual block: a depthwise spatial branch, a depthwise temporal branch, and a grouped spatio-temporal branch, followed by a squeeze-and-excitation channel gate. This reflects the structure of surface code errors, where spatial and temporal syndrome correlations are partially separable. On a unified 4xA100 GPU benchmark, QuantiSpect matches the receptive field of the Accurate baseline at R=13 while using ~2.71x fewer parameters (0.663M vs 1.80M) and ~2.84x fewer per-voxel convolutional MACs. It matches Accurate's circuit-level threshold and accuracy at moderate and large code distances, reduces the logical error rate by up to ~1.85x relative to uncorrelated PyMatching at d=13, p=0.5%, and speeds up the PyMatching decode by up to 3.11x at d=23. We also explored enlarging the receptive field by adding blocks. Even at R=21, the model uses only 1.18M parameters, fewer than both the R=13 Accurate baseline (1.80M) and the R=17 dense model (4.22M), despite its larger receptive field. This expanded variant significantly outperforms the Accurate model, raising the circuit-level threshold to ~0.80% and further reducing the logical error rate. Together, both variants show that a structure-aware factorized design is an effective, parameter-efficient alternative to a dense one for decoding the surface code.