Decoding the surface code with a spatio-temporal transformer

Quantum error correction is a major area of research for building fault-tolerant quantum computers. Recently, machine learning has emerged as a compelling approach to quantum error decoding because of its flexibility and high performance compared to classical methods. In this work, we introduce a spatio-temporal transformer with graph Laplacian positional encodings and factorized latent attention to efficiently deal with the topological structure of the code and the temporal correlations across measurement rounds. We perform experiments with simulated data of the surface code for small distances and demonstrate the potential for our model.