Reinforcement Learning for Quantum Statistical Physics

During the last two decades, Machine Learning (ML) and Artificial Intelligence (AI) tools have created a true paradigm shift and impacted numerous fields and industries. In quantum physics, ML is rapidly gaining popularity and is already being extensively used for variational quantum state representation. Recently, a more ambitious and new research direction is developing where Reinforcement Learning agents could be used to solve quantum statistical problems while improving during the task. This field is still in its infancy and is highly promising to yield efficient and scalable computational tools for physics that would be situated between semi-analytical approximations and brute-force Monte Carlo calculations. In this proposal such tools will be developed for applications in modern quantum many-body physics at finite temperatures. In particular, the goal is to train smart AI agents to sample path integrals that occur in various quantum statistical problems. Important research questions include exploring domain generalization where learned knowledge by the agent can be transferred between tasks. The developed methodology will be applied to challenging systems in condensed matter physics such as many-fermion systems and polaronic systems with memory. Besides providing powerful computational tools, this research on the thrilling synthesis of Reinforcement Learning and quantum statistics will yield new insights and perspectives at the forefront of the current AI explosion in physics.