Reinforcement learning based methods for optimal control and design of quantum systems

Abstract

Finding new and effective methods to optimally design and control quantum systems is crucial for the development of Quantum Technologies. A possible approach is to use Machine Learning, which has recently seen a significant rise in interest and success in many scientific domains, including Physics. Among the various branches of Machine Learning, Reinforcement Learning is a rich and growing field that can be extremely useful to solve problems of planning and control. In this Thesis, we use a Reinforcement Learning approach to address problems of optimal control and design of quantum systems.

We present a general Reinforcement Learning-based strategy to find optimal control protocols for a quantum system. We adapt this approach to cases where the system dynamics cannot be simulated and we cannot perform measurements during the evolution, outperforming traditional numerical optimization in terms of number of experiments required to find an optimum when we only have minimal knowledge on the system. We apply such methodology to Quantum Thermodynamics, reducing the dissipation and irreversibility arising from non-quasistatic transformations of closed quantum systems, and in the context of coherent population transfer, realizing almost perfect transfer between the lower energy states of a three-level lambda system. In the latter case we also combine the proposed Reinforcement Learning based approach with numerical optimization techniques to further improve optimal solutions.

We then shift our attention to quantum system design and use Reinforcement Learning to find optimal qubit chain configurations for end-to-end excitation-transfer. In general, Reinforcement Learning can be deployed for optimal quantum systems design when the problem of designing a physical system for a given task can be translated into a problem of optimal decision making. This is especially useful when we have limited knowledge on the system properties, we need to optimize discrete variables or perform multiple optimizations under different conditions.

Date of Award	Jul 2024
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	The Royal Society
Supervisor	Mauro Paternostro (Supervisor) & Alessandro Ferraro (Supervisor)