Dynamic Bayesian Machine Learning for Quantum Device Control

One of the most import ongoing research avenues is the development of quantum devices sufficient for key technological applications such as quantum computing. As techniques supporting the development of these devices become increasingly refined, and likewise as the demanded scalability increases, the ability of humans to manually execute the required control over these devices is rapidly becoming insufficient. A particularly important example is the case of quantum device tuning, which is a highly challenging task due to the prohibitive cost of evaluations and high dimensionality of the problem as even a single quantum device will have at least a ten-dimensional parameter space Recent research has found remarkable success tackling this problem using machine learning techniques such as deep learning and Bayesian optimisation, developing algorithms capable of efficiently traversing this high-dimensional parameter space without human input. This primarily provides the key benefit of dramatically speeding up the tuning process by taking the human out of the loop, but also allows the execution of improved control over the device and presents the opportunity for enhancing scaling of the tuning process to achieve the ultimate goal of a sufficiently large quantum device (e.g. in the application of quantum computing).
Much of the success of this research is owed to the Bayesian approach taken, which is key to the efficiency of this process due to the use of a probabilistic model allowing principled handling of uncertainty in the system. However, this probabilistic model is based on the idealised physics of the problem such as models of the semiconductors used in these quantum devices. In reality, disorder is always present which can mitigate some of the benefits of the machine learning approach. Accurate characterisation of this disorder and the ability to subsequently inform the tuning algorithms is thus crucial to unlocking the true potential of the application of machine learning in quantum device control. Recent research has once again found success applying machine learning to this problem, particularly the techniques of physics-aware machine learning and Bayesian inference over Gaussian random fields.
A significant problem with the above approach is its static formulation, neglecting the dynamic nature of the disorder of a quantum system. Adaption of these techniques to a dynamic setting is nontrivial and requires significant theoretical development in machine learning itself even before application to this problem. This project aims to tackle exactly this problem through a Bayesian lens which again provides a principled structure for the incorporation of knowledge over time and dynamic decision planning. Particularly, this project aims to make advances in areas such as physics-aware spatiotemporal probabilistic models, fast online Bayesian inference and Bayesian continuous meta-learning which are judged to be key developments necessary to achieving accurate and measurement-efficient characterisation of the disorder of quantum devices. Another potential benefit of this research is in the adjacent problem of scaling the algorithm to many-device quantum systems. Current research again adopts a static formulation of the problem treating each device independently. Bayesian continuous meta-learning may be used to iteratively improve the algorithm between devices, and theory underpinning spatiotemporal probabilistic models may be modifiable to apply to the axis of device rather than time enabling the simultaneous tuning of entangled devices. This project is key to unleashing the potential of quantum technologies by allowing fast tuning and accurate control of quantum devices. This project falls within the EPSRC Quantum technologies research area.The project is in collaboration with Georgios Katsaros (IST Austria) and Dominik Zumbuhl (University of Basel), both of whom are providing physical quantum devices in support of experiments.