Abstract
Designing experiments for high-dimensional quantum entanglement is a complex task with an astronomical search space. We present an artificial intelligence agent based on the Projective Simulation framework that autonomously designs quantum optics experiments. The agent learns to navigate the experimental space, rediscovering sophisticated techniques like parity sorters and inventing new "macro-actions" to efficiently create target entangled states. Our results show that this active learning approach significantly outperforms traditional automated random search methods in both discovery rate and solution optimization.
Introduction
The creation of high-dimensional ($d > 2$) and multipartite ($n > 2$) entangled states is a cornerstone of modern quantum information science, essential for advanced quantum computing and secure communication protocols. However, the design of optical setups to achieve these states is notoriously difficult. Human researchers often rely on intuition or trial-and-error to arrange components like beam splitters, Dove prisms, and holograms. Previous computational approaches, such as the MELVIN algorithm, utilized random searches and handcrafted rules but lacked the ability to learn the underlying structure of the experimental domain. This study investigates whether a reinforcement learning (RL) agent can develop its own "intuition" to propose novel experimental configurations.
Methods: Projective Simulation and Action Composition
The researchers utilized Projective Simulation (PS), a physics-inspired RL framework. Unlike standard neural networks, PS uses a memory network of "clips"—nodes representing percepts (the state of the optical table) and actions (placing a component).
Key features of the PS agent include:
- Action Composition: Successful sequences of actions are bundled into "macro-actions" or "gadgets," allowing the agent to place complex sub-assemblies in a single step.
- Clip Deletion: To manage a state space exceeding 0.5 billion configurations, the agent prunes nodes and edges that do not contribute to rewards, maintaining computational efficiency.
- Simulated Environment: The agent operates on a virtual optical table with 30 basic elements (beam splitters, Dove prisms, holograms, mirrors) across four optical arms. A simulated analyzer evaluates the resulting quantum state against target Schmidt-Rank Vectors (SRV).
Results
The PS agent demonstrated a superior ability to discover unique and optimized experimental setups compared to random search baselines.
Transfer Learning and Rediscovery
A significant finding was the benefit of "stepping stone" learning. Agents first trained on simpler (3,3,2) entangled states were much faster at finding complex (3,3,3) states than those starting from scratch. Furthermore, the agent autonomously "rediscovered" the parity sorter, a sophisticated interferometric setup used by human physicists to sort photons by orbital angular momentum, identifying it as a highly effective "gadget" for reaching its goals.
Performance Benchmarks
As shown in Figure 6b, the PS agent with action composition found approximately 800 unique, interesting experiments for a maximum length of 12, whereas the random search baseline found only 200. The agent also learned to optimize for the shortest possible setups, consistently reaching the theoretical minimum length for various target states. In complex tasks where random search success is negligible, the trained PS agent achieved a success probability of 50% per trial.
Discussion and Conclusion
This work demonstrates that RL agents can navigate the "maze" of quantum experimental design by identifying structural clusters of successful configurations. While the current study is limited to simulation and does not account for real-world noise like mechanical vibrations, it paves the way for "smart laboratories." Future iterations could see AI agents controlling robotic arms on physical optical tables to discover new physics in real-time, scaling these techniques to even larger multipartite systems.
Summary updated