TY - BOOK

T1 - Machine learning in quantum mechanical and optical systems

AU - Schmiesing, Viktoria-Sophie

PY - 2025/1/29

Y1 - 2025/1/29

N2 - In recent years, machine learning (ML) has become increasingly important across science, industry, and daily life. Simultaneously, quantum computing has emerged as a field with the potential to revolutionize computation. This thesis explores the application of ML to quantum and optical systems. We present two main results: the proposal of a quantum recurrent neural network (QRNN) architecture and the use of reinforcement learning (RL) for experimental fiber coupling, a common task in quantum labs. When data is quantum in nature, quantum ML techniques may be better suited than classical approaches. One such technique is the feed-forward dissipative quantum neural network (DQNN), which learns general quantum channels from independent and identically distributed data. However, many quantum tasks involve sequential data, such as learning quantum state evolution under time-dependent Hamiltonians or interacting with quantum environments. In classical ML, such tasks can, e.g., be handled by recurrent neural networks (RNNs). This thesis proposes a fully quantum RNN structure designed for qudits, named dissipative quantum recurrent neural network (DQRNN). Extending the DQNN framework to the recurrent case, DQRNNs can approximate general causal quantum automata. We present both quantum and classical training algorithms for DQRNNs, showing that the resource requirements scale with the width of the underlying DQNN but not with its depth. Numerical results show that DQRNNs solve memory-dependent tasks beyond the capacity of DQNNs, generalizing well from limited data. One promising future application of DQRNNs is model-based RL in quantum environments. Although RL is inherently well-suited for control tasks, RL agents for applications in optical experiments have mostly been trained in simulation. Taking the example of fiber coupling, we show that it is feasible to apply RL directly in experiments. This saves us the time of extensive system and noise modeling. Still, intermediate challenges needed to be overcome such as time-consuming training, noisy actions, and partial observability. For shorter training times, we use a simple virtual testbed for environment tuning and algorithm selection. We demonstrate that an RL agent can learn to overcome noisy actions and partial observability. In four days of training time directly in the experimental setup, using sample-efficient algorithms such as truncated quantile critics (TQC) and soft actor-critic (SAC), it learns to achieve coupling efficiencies comparable to human experts. This thesis takes key steps toward integrating machine learning in quantum control, introducing DQRNNs that could serve as a powerful tool for modeling quantum environments and highlighting the role of RL in experimental physics. By showing how RL can be applied successfully directly in an optical experiment using the example of fiber coupling, this work paves the way for applying RL to more intricate quantum systems.

AB - In recent years, machine learning (ML) has become increasingly important across science, industry, and daily life. Simultaneously, quantum computing has emerged as a field with the potential to revolutionize computation. This thesis explores the application of ML to quantum and optical systems. We present two main results: the proposal of a quantum recurrent neural network (QRNN) architecture and the use of reinforcement learning (RL) for experimental fiber coupling, a common task in quantum labs. When data is quantum in nature, quantum ML techniques may be better suited than classical approaches. One such technique is the feed-forward dissipative quantum neural network (DQNN), which learns general quantum channels from independent and identically distributed data. However, many quantum tasks involve sequential data, such as learning quantum state evolution under time-dependent Hamiltonians or interacting with quantum environments. In classical ML, such tasks can, e.g., be handled by recurrent neural networks (RNNs). This thesis proposes a fully quantum RNN structure designed for qudits, named dissipative quantum recurrent neural network (DQRNN). Extending the DQNN framework to the recurrent case, DQRNNs can approximate general causal quantum automata. We present both quantum and classical training algorithms for DQRNNs, showing that the resource requirements scale with the width of the underlying DQNN but not with its depth. Numerical results show that DQRNNs solve memory-dependent tasks beyond the capacity of DQNNs, generalizing well from limited data. One promising future application of DQRNNs is model-based RL in quantum environments. Although RL is inherently well-suited for control tasks, RL agents for applications in optical experiments have mostly been trained in simulation. Taking the example of fiber coupling, we show that it is feasible to apply RL directly in experiments. This saves us the time of extensive system and noise modeling. Still, intermediate challenges needed to be overcome such as time-consuming training, noisy actions, and partial observability. For shorter training times, we use a simple virtual testbed for environment tuning and algorithm selection. We demonstrate that an RL agent can learn to overcome noisy actions and partial observability. In four days of training time directly in the experimental setup, using sample-efficient algorithms such as truncated quantile critics (TQC) and soft actor-critic (SAC), it learns to achieve coupling efficiencies comparable to human experts. This thesis takes key steps toward integrating machine learning in quantum control, introducing DQRNNs that could serve as a powerful tool for modeling quantum environments and highlighting the role of RL in experimental physics. By showing how RL can be applied successfully directly in an optical experiment using the example of fiber coupling, this work paves the way for applying RL to more intricate quantum systems.

U2 - 10.15488/18472

DO - 10.15488/18472

M3 - Doctoral thesis

CY - Hannover

ER -