Quantum Control And Quantum Information Processing Via Deep Reinforcement Learning

Statistics-based machine learning is the science of using big data to learn from and make predictions.Data collection and analysis are a characteristic of modern science,including physics.Data analysis is equally important in different fields within physics,such as particle physics experiments,astronomical observations,cosmology,condensed matter physics,biophysics,and quantum computing,which require analysis of simulation and experimental data.On the other hand,quantum informatics,as another representative of emerging science,has a high probability of having a significant impact on our future society.As two independent fundamental research fields,quantum information and machine learning have their research problems and challenges and have been heavily studied.However,researchers have found that developments in these two fields can largely reinforce and benefit each other in recent studies.Quantum control is a crucial technology for achieving many quantum information tasks.However,manipulable quantum resources are fragile and easily corrupted by the surrounding environmental noise,and there is no reasonable control theory to guide the control of high-dimensional quantum systems.From this aspect,it is necessary to study the quantum dynamics of high-dimensional open quantum systems to reduce the interference of environmental noise.On the other hand,the quantum phase transition is a fundamental phenomenon in condensed matter physics,which has attracted the attention of many researchers.How to reveal the critical properties of quantum many-body systems is a significant task.The traditional research methods focus on the order parameter and symmetry breaking.Recent advances in quantum information theory have provided a new perspective for studying quantum phase transitions,especially in the absence of known order covariance,and concepts in quantum information such as quantum correlation and fidelity have been successfully used to reveal the nature of quantum phase transitions.Reinforcement learning is a subfield of machine learning that focuses on how intelligence interacts with their environment to give better strategies for specific tasks.Research in recent years has demonstrated the advantages of deep reinforcement learning for numerical computation in quantum physics,including quantum state preparation,quantum line design,and fault-tolerant quantum computation.The research is done in this thesis mainly includes the following:1.A quantum control model for preparing quantum gates using discrete sequential control is investigated,and a deep reinforcement learning algorithm is used to perform optimal control for preparing target quantum logic gates.A general framework is developed to link quantum dynamics with optimal control.Moreover,numerical simulation results of three different traditional algorithms are computed to compare with deep reinforcement learning.It is demonstrated that the framework can effectively learn and control the system to reach the target quantum gate approximately optimally.In addition,it is found that the reinforcement learning algorithm can effectively identify and reduce the effect of local minima in the policy space.2.A novel high-dimensional quantum control model with discrete sequential control is investigated for preparing target quantum states in noise environments.A framework is developed to link the quantum dynamics control of this model with reinforcement learning algorithms.The results show that the deep reinforcement learning algorithm can find better control strategies in up to 10-dimensional Hilbert space when the environment has disturbances(e.g.,dephasing and energy decay).3.A method to optimize the evolutionary path of quantum imaginary time evolutionary algorithms with deep reinforcement learning is investigated to minimize the algorithm errors.A general framework is developed to link deep reinforcement learning with quantum optimization algorithms.Numerical simulation results of the optimized quantum algorithm under two different models are computed.Moreover,experimental results verify the scheme’s effectiveness using a nuclear magnetic resonance quantum computer.The results provide a novel approach to studying quantum optimization algorithms.4.The relationship between local entanglement structures and quantum phase transitions is investigated.A novel approach to identifying quantum phase transitions in Ising spin chains is created with deep reinforcement learning to design the single-qubit local disentanglement quantum circuit.In this thesis,we first propose a deep reinforcement learning framework for designing variational quantum circuits.This framework is then applied to the one-dimensional transverse field Ising model and the onedimensional XXZ model to obtain specific quantum circuits.The disentanglement quantum circuits designed by reinforcement learning can perfectly distinguish the quantum phase transitions of the transverse field Ising model and the XXZ model.The results provide a new perspective for studying quantum entanglement structures and quantum phase transitions.