Evaluating Learned Decoders under Compound Noise Models in Surface Codes

Hiroshi Tanaka, Syed Ali, Muhammad Salik Ali Khan

The practical realization of fault-tolerant quantum computing critically hinges on the capacity of quantum error correction (QEC) schemes to robustly mitigate physical noise. Among the foremost candidates, the surface code is distinguished by its high threshold and local stabilizer checks. In recent years, neural network-based decoders, particularly those leveraging deep reinforcement learning (RL) and convolutional architectures, have demonstrated competitive performance against algorithmic staples such as Minimum Weight Perfect Matching (MWPM) and Union-Find. However, the vast majority of these learned decoders are trained and evaluated under idealized phenomenological or circuit-level depolarizing noise models. Real-world quantum processors exhibit complex, compound noise profiles encompassing crosstalk, leakage, non-Markovian temporal drift, and spatially correlated errors. In this work, we empirically evaluate the robustness of state-of-the-art neural decoders against highly structured compound noise models. We introduce a comprehensive stress-testing methodology tailored to quantify the degradation of the logical error rate when decoders are subjected to out-of-distribution noise characteristics. Our empirical results reveal that while neural decoders can overfit to their training noise profile, strategically designed training curricula and architecture regularizations enable significant resilience. We provide extensive numerical benchmarks validating that learned decoders can maintain near-optimal pseudothresholds across a broad spectrum of shifting noise conditions, highlighting their readiness for real-world superconducting qubit deployments.

Savvas Varsamopoulos, Koen Bertels, Carmen G. Almudéver

Ali Syed, Tanaka Hiroshi, Khan Muhammad

The practical realization of fault-tolerant quantum computing critically hinges on the capacity of quantum error correction (QEC) schemes to robustly mitigate physical noise. Among the foremost candidates, the surface code is distinguished by its high threshold and local stabilizer checks. In recent years, neural network-based decoders, particularly those leveraging deep reinforcement learning (RL) and convolutional architectures, have demonstrated competitive performance against algorithmic staples such as Minimum Weight Perfect Matching (MWPM) and Union-Find. However, the vast majority of these learned decoders are trained and evaluated under idealized phenomenological or circuit-level depolarizing noise models. Real-world quantum processors exhibit complex, compound noise profiles encompassing crosstalk, leakage, non-Markovian temporal drift, and spatially correlated errors. In this work, we empirically evaluate the robustness of state-of-the-art neural decoders against highly structured compound noise models. We introduce a comprehensive stress-testing methodology tailored to quantify the degradation of the logical error rate when decoders are subjected to out-of-distribution noise characteristics. Our empirical results reveal that while neural decoders can overfit to their training noise profile, strategically designed training curricula and architecture regularizations enable significant resilience. We provide extensive numerical benchmarks validating that learned decoders can maintain near-optimal pseudothresholds across a broad spectrum of shifting noise conditions, highlighting their readiness for real-world superconducting qubit deployments.

Claire E. Badger

Shay J. Manor, Leila S. Erhili, Yassine Jebbouri

Quantum error-correction decoders are typically benchmarked against synthetic circuit-level noise, under the assumption that a decoder's ranking under such noise transfers to hardware and improves as the noise model becomes more realistic. The Willow processor, the first to operate below the surface-code threshold, allows us to test this assumption. We rank a panel of six decoders using a four-rung ladder of noise models with increasing fidelity, evaluated against real data across three code distances, two bases, and fifteen round counts. Rank agreement with hardware appears once the noise model gives each operation type its own error rate. Calibrating the model to the device improves absolute error rates but not rank agreement. We additionally provide the first independent evaluation of NVIDIA's Ising pre-decoder on hardware, at code distances below its training receptive field and via a mapping onto the lattice on which it was trained. Under these conditions, it holds no accuracy-latency advantage: another panel decoder matches or improves on it in both per-cycle error rate and decode latency in 278 of the 280 evaluations. We release the full pipeline and the per-shot outcome of every evaluation, so future decoders and devices can be compared.

A Hussain, Priyanka Rani, Jayapal Lande +3 more