Large-scale quantum computation requires a reliable assessment of the main sources of error in the implemented quantum gates. To this aim, we provide a learning-based framework that extracts the contribution of each physical error source to the infidelity of a series of gates, together with an uncertainty estimate for every contribution. Crucially, the uncertainties provided by the underlying Gaussian Process Regression serve as a diagnostic of the assumed noise model itself: if the physical model does not fully explain the experimental observations, the framework signals this instead of returning a falsely precise error budget. We demonstrate the proposed budget learning procedure on a pair of qubits in a transmon array architecture, where we compute the error budgets of the native single- and two-qubit gates. For this purpose, we have also developed an advanced noise model based on the first-principles understanding of the error processes in current superconducting processors, including complex (e.g. non-Markovian) error sources. While the single-qubit budgets are reconstructed with high confidence, the same diagnostic reveals that even this detailed noise model does not fully capture the error dynamics of a state-of-the-art two-qubit gate --- quantitative evidence that uncertainty-aware error budgeting is necessary for the characterization of current quantum hardware.
High fidelity quantum operations are key to enabling fault-tolerant quantum computation. Superconducting quantum processors have demonstrated high-fidelity operations, but on larger devices there is commonly a broad distribution of qualities, with the low-performing tail affecting near-term performance and applications. Here we present an error budgeting procedure for the native two-qubit operation on a 32-qubit superconducting-qubit-based quantum computer, the OQC Toshiko gen-1 system. We estimate the prevalence of different forms of error such as coherent error and control qubit leakage, then apply error suppression strategies based on the most significant sources of error, making use of pulse-shaping and additional compensating gates. These techniques require no additional hardware overhead and little additional calibration, making them suitable for routine adoption. An average reduction of 3.7x in error rate for two qubit operations is shown across a chain of 16 qubits, with the median error rate improving from 4.6
% to 1.2
% as measured by interleaved randomized benchmarking. The largest improvements are seen on previously under-performing qubit pairs, demonstrating the importance of practical error suppression in reducing the low-performing tail of gate qualities and achieving consistently good performance across a device.
We demonstrate new experimental procedures for measuring small errors in a superconducting quantum bit (qubit). By carefully separating out gate and measurement errors, we construct a complete error budget and demonstrate single qubit gate fidelities of 0.98, limited by energy relaxation. We also introduce a new metrology tool-- Ramsey interference error filter-that can measure the occupation probability of the state |2> which is outside the computational basis, down to 10{-4}, thereby confirming that our quantum system stays within the qubit manifold during single qubit logic operations.