Learning correlated noise in a 39-qubit quantum processor

Building error-corrected quantum computers relies crucially on measuring and modeling noise on candidate devices. In particular, optimal error correction requires knowing the noise that occurs in the device as it executes the circuits required for error correction. As devices increase in size we will become more reliant on efficient models of this noise. However, such models must still retain the information required to optimize the algorithms used for error correction. Here we propose a method of extracting detailed information of the noise in a device running syndrome extraction circuits. We introduce and execute an experiment on a superconducting device using 39 of its qubits in a surface code doing repeated rounds of syndrome extraction, but omitting the mid-circuit measurement and reset. We show how to extract from the 20 data qubits the information needed to build noise models of various sophistication in the form of graphical models. These models give efficient descriptions of noise in large-scale devices and are designed to illuminate the effectiveness of error correction against correlated noise. Our estimates are furthermore precise: we learn a consistent global distribution where all one- and two-qubit error rates are known to a relative error of 0.1%. By extrapolating our experimentally learned noise models towards lower error rates, we demonstrate that accurate correlated noise models are increasingly important for successfully predicting sub-threshold behavior in quantum error correction experiments.