Optimizing quantum dynamics in the age of machine learning
File(s)
Author(s)
Sauvage, Frederic
Type
Thesis
Abstract
Optimizing the dynamics of quantum systems enables the design of high precision experiments and the development of quantum technologies. To date, such optimizations have been predominantly performed based on theoretical models and numerical simulations; given the size and intricacy of the systems that can now be controlled, this approach is reaching its limits. Optimizing dynamics based solely on experimental data provides a mean to exceed these limitations. However, the probabilistic nature of quantum measurements, combined with the relatively low repetition rates and high noise levels in many current experiments, render these optimizations uniquely difficult. Inspired by recent developments in the field of machine learning, we develop new methods for the efficient optimization of quantum dynamics in experimental situations.
First, we establish Bayesian optimization as a methodology well suited for these optimization problems. After thoroughly assessing its benefit on a paradigm problem of quantum optimal control, we refine the original framework to take into consideration the statistical features of quantum measurements. This allows to maximize the utility of each measurement data and results in enhanced convergence of the framework.
Going further, we investigate an aspect of optimization often ignored, namely the choice of the figure of merit. This figure, which for any optimization problem is identified as the quantity to be maximized, is not unique, thus leaving room for its refinement. After establishing criteria for adequate figures of merit, we show that improved figures, compared to several canonical ones, can be designed.
Finally, to fully reap the benefits of experimental optimizations, it is often desirable that not only one but many related optimizations are performed concurrently. To this intent, we propose two novel frameworks which are found to yield substantial improvements compared to existing methodologies.
First, we establish Bayesian optimization as a methodology well suited for these optimization problems. After thoroughly assessing its benefit on a paradigm problem of quantum optimal control, we refine the original framework to take into consideration the statistical features of quantum measurements. This allows to maximize the utility of each measurement data and results in enhanced convergence of the framework.
Going further, we investigate an aspect of optimization often ignored, namely the choice of the figure of merit. This figure, which for any optimization problem is identified as the quantity to be maximized, is not unique, thus leaving room for its refinement. After establishing criteria for adequate figures of merit, we show that improved figures, compared to several canonical ones, can be designed.
Finally, to fully reap the benefits of experimental optimizations, it is often desirable that not only one but many related optimizations are performed concurrently. To this intent, we propose two novel frameworks which are found to yield substantial improvements compared to existing methodologies.
Version
Open Access
Date Issued
2021-08
Date Awarded
2021-11
Copyright Statement
Creative Commons Attribution-Non Commercial 4.0 International Licence
License URL
Advisor
Mintert, Florian
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/P510257/1
Publisher Department
Physics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)