Human intelligence is associated with the ability to perform complex reasoning, find creative solutions to problems, and discover and generalize patterns. Ambitiously, the field of artificial intelligence (AI) aims to enable computer systems to acquire such abilities, and has experienced remarkable progress in recent years. Leveraging these advances for scientific research has the potential to accelerate progress and achieve milestones previously unreachable. In this thesis, we develop and adapt new tools based on AI and numerical optimization to address several challenges in quantum information processing, particularly in quantum control and the design of quantum experiments. Among AI methods, reinforcement learning (RL) provides a framework for discovering optimal strategies for controlling quantum systems without requiring prior knowledge of the system. However, it had been an outstanding challenge to implement and train such an RL-based controller capable of interacting with a quantum experiment in real-time, so on time scales much shorter than the system's coherence time. In our work, we overcome this challenge by deploying a sub-microsecond-latency neural network in a superconducting circuit experiment. Using RL, we train the network exclusively from measurements and show that it can efficiently initialize a superconducting transmon. Next, we experimentally test an improved protocol for the quantum control of a microwave cavity, which promises to completely suppress coherent errors in the so-called SNAP quantum gate. We apply this scheme in a circuit quantum electrodynamics experiment and demonstrate that it reduces the dominant error of the SNAP gate by 53% compared to the original protocol in realistic scenarios. A central challenge and opportunity is to harness AI methods to acquire new scientific knowledge. This is the ambition of the field of artificial scientific discovery. In this spirit, we develop the algorithm AutoScatter to automatically discover new design schemes for signal-processing devices. AutoScatter optimizes the discrete and continuous system parameters to achieve the desired target behavior with the minimum required resources. By using an abstract graph representation, the discovered setups can be interpreted easily, leading to multiple generalizations. Our design concepts are transferable and can be implemented on a variety of hardware platforms, including photonic, microwave, and optomechanical systems. In the last part of this thesis, we extend AutoScatter to the discovery of systems with a periodic lattice structure, leading to new amplifier, isolator, and frequency-demultiplexer schemes. To do so, we base our automated discovery on a transfer-matrix approach, offering a new perspective on non-Hermitian topological systems.
Jonas Landgraf (Thu,) studied this question.