Two-Stage Neural Residual Decoding for FastSurface-Code Quantum Error Correction

Technical Summary: Two-Stage Neural Residual Decoding for Fast Surface-Code Quantum Error Correction

Problem Statement
Fault-tolerant quantum computation requires fast and reliable decoding of surface codes. While Minimum-Weight Perfect Matching (MWPM) is a standard, reliable decoder for surface codes, its computational cost increases significantly as code distance, circuit depth, or physical error rates grow, making real-time operation challenging. Conversely, neural-network decoders offer speed but often suffer from poor transferability; large models typically require costly retraining when code distances or noise models change. Furthermore, deploying a neural decoder directly on a full detector graph introduces complexities regarding detector ordering, boundary conditions, and logical observable conventions, making it difficult to replace mature graph-based decoders entirely.

Methodology
The authors propose a two-stage hybrid decoding framework that combines a pretrained Graph Neural Network (GNN) for local residual pre-decoding with a global MWPM backend for final cleanup.

1. Architecture and Workflow:

   • Stage 1 (Neural Pre-decoding): A pretrained GNN receives a detector graph generated by an established decoding backend. Instead of predicting a full correction, the GNN acts as a "residual pre-decoder." It assigns confidence scores to local correction candidates defined by the backend.
   • Confidence Filtering: Only candidates exceeding a fixed confidence threshold $\tau$ are selected. Uncertain regions are left untouched, preserving the residual syndrome for the global stage.
   • Stage 2 (Global Cleanup): The residual syndrome, updated by the accepted local corrections, is passed to an NVIDIA-accelerated MWPM backend. This backend handles the final logical decision, boundary connections, and non-local correlations using standard decoding conventions.

2. Parallel-Window Inference:
   To accelerate the neural stage, the authors implement a two-layer staggered parallel-window strategy.

   • Layer A: Evaluates a set of non-overlapping local windows in parallel.
   • Layer B: Evaluates spatially offset windows after the Layer A update, allowing the GNN to inspect residual patterns that cross Layer A boundaries.
   • This approach avoids serial scanning of overlapping neighborhoods and enables batched processing of local subgraphs.

3. Training Strategy:
   The GNN is trained offline using supervised learning on detector graphs and labels generated by the official backend. The network learns to identify high-confidence local residual corrections using binary cross-entropy loss. Once trained, the GNN parameters are frozen, and the model does not require retraining for inference, though separate models are trained for specific code distances and noise models.

Key Contributions

• Hybrid Decoding Strategy: The paper introduces a method where the neural network augments, rather than replaces, the MWPM decoder. The GNN focuses on frequent, reliable local patterns, while MWPM handles the complex global structure and boundary conditions.
• Preservation of Backend Conventions: By relying on the backend for graph construction, detector ordering, and boundary treatment, the method avoids the need for the neural model to reconstruct these quantities, ensuring consistency with established decoding protocols.
• Parallel-Window Acceleration: The implementation of staggered, non-overlapping local windows allows for efficient batched inference, significantly reducing latency compared to serial processing.

Experimental Results
The method was evaluated on code distances $d = 3, 5, 7, 9$ across physical error rates $p = 0.003$ to $0.021$ (and up to $0.028$ for Logical Error Rate analysis) using batches of 32,768 samples.

• Speedup:

  • At $d=3$, the complete method achieved a 3.48x to 4.38x speedup over Pure MWPM for error rates between $p=0.003$ and $0.021$.
  • Ablation studies confirmed that the majority of this speedup comes from the parallel GNN inference over fixed windows, rather than just the reduction in residual syndrome size (the variant without GNN achieved only 1.06x–1.30x speedup).
  • Across 160 configurations involving four different circuit-level noise models (SI1000, Z-biased, measurement-dominated, and calibration-drifted), the method achieved speedups ranging from 2.25x to 5.41x.

• Logical Failure Rate (LER):

  • The hybrid method reduced the Logical Failure Rate relative to Pure MWPM in most tested scenarios. For example, at $p=0.007$, LER reductions were 18.01% ($d=5$), 9.71% ($d=7$), and 13.32% ($d=9$).
  • The improvement in LER was noise-model dependent. The method showed significant LER benefits under calibration-drifted and measurement-dominated noise but showed mixed results for SI1000 and strongly Z-biased noise, though it consistently maintained or improved speed.

Significance and Claims
The authors claim that this work provides a practical stage for surface-code decoding that accelerates the decoding pace without sacrificing, and in some cases improving, logical decoding performance.

• Scalability: The approach addresses the computational bottleneck of MWPM in real-time operations by offloading local pattern recognition to efficient, parallelized neural inference.
• Robustness: By retaining the trusted MWPM backend for global decisions, the system avoids the pitfalls of neural decoders that might fail on unseen global configurations or violate logical observable conventions.
• Extensibility: The framework is shown to be extensible to various circuit-level noise models, provided that separate GNNs are trained for specific code distances and noise configurations. The authors note that while the current implementation requires separate training for each setting, the architecture supports transfer across configurations in future work.

The paper concludes that this two-stage approach represents a viable operating point for hybrid decoders, effectively balancing the speed of neural inference with the reliability of graph-based matching.