Machine Learning Decoders Quantum Error Correction

Short answer: Error correction needs a decoder that turns noisy syndrome measurements into a correction, and several research groups have trained neural networks to do this for the surface code. A 2024 Nature paper reports a transformer-based decoder that beat other state-of-the-art decoders on data from a Google superconducting processor, a 2022 paper reports a scalable convolution-style decoder, and a 2025 paper compares a neural decoder with minimum-weight perfect matching. All of this is research on small to moderate code sizes.

The decoding problem

The 2024 Nature paper explains that building a large quantum computer needs strategies to correct errors that inevitably arise, and that codes encode logical information redundantly across many physical qubits. A key challenge is accurately decoding noisy syndrome information from the redundancy checks to recover the encoded logical information.

A transformer-based decoder

The Nature authors develop a recurrent, transformer-based neural network, which they call AlphaQubit, that learns to decode the surface code. They report it outperforms other state-of-the-art decoders on real data from Google's Sycamore processor for distance-3 and distance-5 surface codes. On distances up to 11, it keeps its advantage on simulated data with realistic noise including cross-talk and leakage, using soft readouts and leakage information. Training has two stages: pretraining on approximate synthetic data, then adapting to the more complex, unknown real error distribution with a limited budget of experimental samples. The authors present this as machine learning going beyond human-designed algorithms by learning directly from data.

A scalable neural decoder

A 2022 Physical Review Letters paper from the University of Cologne presents a neural decoder for a family of stabilizer codes under depolarizing noise and syndrome measurement errors. The authors say it scales to tens of thousands of qubits, unlike other recent machine-learning decoders, and decodes faster than the state-of-the-art union-find decoder over a wide range of error rates down to 1 percent. The key idea is to decode error syndromes on small scales by sliding a preprocessing window over the code, like a convolutional network. They report that this lowers the effective error rate and, by detecting correlations, raises the error threshold above that of conventional decoding.

Comparison with matching

A 2025 Physical Review Research paper studies a neural decoder on small-distance surface codes using simulated data and experimental data from Google Quantum AI's 2023 Nature experiment on a transmon processor. It finds the network typically beats minimum-weight perfect matching because it handles errors that cause several correlated syndrome defects better, approaches the performance of a maximum-likelihood decoder on the experimental data, and can use the soft information in analog readout. The authors say neural decoders need no prior information about physical error rates, which makes them adaptable, and that their performance and efficiency suit near-term quantum memory demonstrations.

What this does not tell you

The comparisons are against the specific decoders and data sets named in each paper. The experimental results are for small distances, and larger distances rely on simulation. Decoding speed on real-time hardware is a separate question the papers do not settle.

FAQ

What does a decoder do?

It turns noisy syndrome measurements into a most likely correction for the encoded logical information.

Can neural networks beat standard decoders?

The 2024 and 2025 papers report this on Google processor data for small surface codes.

Do they need known error rates?

The 2025 paper says neural decoders need no prior information about physical error rates.

Do they scale?

The 2022 paper reports a design scalable to tens of thousands of qubits in simulation.

Is it settled?

No. These are research results, with larger distances evaluated only on simulated data.