Condensed Matter Seminar Series

Natalia Ares

Natalia Ares Group, Materials Department, University of Oxford, UK

Quantum device measurement and tuning using machine learning

Fulfilling the promise of quantum technologies requires the ability to operate many devices; fault-tolerant factorization using a surface code will require ~108 physical qubits. A long-term approach, based on the success of integrated circuits, is to use solid-state qubit devices. A major obstacle to creating large circuits in this platform is device variability. It is very time consuming to fully characterize and tune each of these devices and this task will rapidly become intractable for humans without the aid of automation.

I will present efficient measurements on a quantum dot performed in real time by a machine learning algorithm. This algorithm employs a probabilistic deep-generative model, capable of generating multiple full-resolution reconstructions from scattered partial measurements. Information theory is used to select the most informative measurements to perform next (Fig.1). The algorithm outperforms standard grid scan techniques in different measurement configurations, reducing the number of measurements required by up to 4 times.

I will also show the use of Bayesian optimisation to tune quantum dot devices. By generating a score function, we can efficiently navigate a multi-dimensional parameter space. We tune the device to the single-electron tunnelling regime with no previous knowledge of the device characteristics in less than a thousandth part of the time that it requires manually.

Figure 1 

Figure 1. With training and sparse initial measurements, the machine computes an information gain map (i), on the basis of which it decides the next measurements to be performed (ii), and it updates the reconstructions, which are guesses of how the stability diagram might look like (iii). The reconstructions resemble more the full-resolution measurement (iv) with each measurement. White regions correspond to constant current, which carry little information and which the algorithm wisely avoids.