QuantiSpect: A Structure-Aware Lightweight 3D CNN Pre-Decoder for Scalable Surface Code Quantum Error Correction

As quantum processors scale, real-time decoding of surface code syndromes becomes a critical bottleneck. While AI-based pre-decoders can accelerate this process by locally correcting errors before passing them to a global decoder, existing architectures rely on dense 3D convolutions that are computationally expensive and parameter-heavy. This paper asks whether a more efficient, structure-aware neural network architecture can match the performance of these dense models while reducing computational overhead.

The authors introduce QuantiSpect, a lightweight 3D CNN designed to replace the dense backbones used in previous neural pre-decoders. The core innovation is the use of factorized residual blocks. Instead of applying dense 3D convolutions, each block uses three parallel branches: a depthwise spatial branch, a depthwise temporal branch, and a grouped spatio-temporal branch. This design explicitly mirrors the physical reality of surface code errors, where spatial lattice patterns and temporal syndrome correlations are partially separable. The architecture also incorporates squeeze-and-excitation gates to weight channel importance and uses GroupNorm for stability with sparse syndrome inputs.

QuantiSpect demonstrates significant efficiency gains without sacrificing accuracy. At a code distance of d=13, the model matches the circuit-level threshold and decoding accuracy of the 'Accurate' baseline while requiring approximately 2.71x fewer parameters and 2.84x fewer per-voxel operations. Furthermore, the authors show that the modular design allows for easy scaling of the receptive field. An expanded version of the model (R=21) uses fewer parameters than the baseline R=13 model yet achieves superior performance, raising the circuit-level error threshold to approximately 0.80% and further reducing logical error rates.