A key requirement for fast iterative R&D of quantum devices is the ability to perform rapid, in-depth characterisation to guide the design of next-generation devices. The Qruise Rapid Automated Bring-up framework lets you connect characterisation experiments to a model database, define sequences and dependencies between these experiments and then trigger them to run in an automated batch mode or in an interactive debug mode. Use standard experiments from our library or bring your own. A very comprehensive storage, logging, monitoring and analysis framework lets you review past runs, track high level system parameters and easily view error logs or quickly jump into unexpectedly failing experiments. All these seamlessly connect with the quantum hardware through APIs from control electronics software. Additionally, QruiseOS create a Qiskit quantum cloud service with your experiments to submit jobs and monitor metrics on a live dashboard with all the necessary housekeeping tasks taken care of.
At the heart of the Model Learning process is a fully differentiable Digital Twin that accurately models the pulse level quantum dynamics of the QPU and all associated controls. This Digital Twin is parameterized and accurately learns system parameters from arbitrary experimental data. Using back-propagation and a gradient based algorithm, the Model Learning process iteratively reduces the statistical distance between the output of the Digital Twin and the real QPU stack, thus obtaining a close match between the two.
The key advantage of a Predictive Model is its ability to provide actionable insights when it comes to identifying high-value efforts in the R&D of quantum technology devices. This is made possible through the generation of an Error Budget – a detailed break-down of the contributions of various device and control imperfections to the bottom-line performance metric. Once an Error Budget is available, the Digital Twin can be used for simulating a broad range of What-If situations to understand how changes in various system parameters will ultimately improve or degrade the performance metric.
A typical observation when plotting iSWAP-type interactions between 2 coupled qubits is the presence of distorted and asymmetric Chevron patterns arising primarily from non-ideal flux pulses. We demonstrate the simultaneous learning of both QPU parameters (e.g., coupling strength) and arbitrary control Transfer Functions (FIR + IIR) from experimental data using our Model Learning algorithms with the Digital Twin. These parameters are then validated by reproducing the distorted bipolar SNZ-type Chevrons (bottom left). Additionally, the learned Transfer Function is then inverted to obtain the pre-distortion necessary to generate pre-compensated flux pulses which then produce Chevrons with almost no asymmetry or distortions (bottom right). During the entire training process, the Digital Twin is never exposed to the bipolar data, highlighting the predictive power and data efficiency of our algorithms.