Decoding surface-code experiments with a recurrent neural network decoder

Quantum error correction offers a way to reach the low error rates needed to perform useful computations at the expense of an increase in the qubit overhead. A more accurate decoder leads to better logical performance, which can substantially reduce this overhead. Neural network decoders are particularly promising since they do not require any prior information about the physical noise.

In this talk, we explore the performance of such a decoder on both simulated data and the data from the recent surface code experiment performed in [Google Quantum AI, Nature 614, 676–681 (2023)]. This decoder typically outperforms perfect-matching decoders as it can better deal with errors leading to multiple syndrome defects, such as Y errors. When applied to the experimental data, we show that this decoder achieves logical error rates approaching those obtained using an approximate maximum-likelihood decoder. We demonstrate the flexibility of such a decoder by providing the soft information available in the analog readout of transmon qubits and show that this can further decrease the logical error rate.