A collaborative research study co-authored by Prof. Mario Rüttgers has been published in Nature. The study presents HydroGym, a computational platform that combines reinforcement learning (RL) and computational fluid dynamics (CFD) for the development of flow-control strategies.
A particularly striking result of the study is that learned flow-control strategies can transfer to substantially different flow environments without retraining. The authors demonstrate this transfer across different flow configurations, including from a simple channel flow to a three-dimensional aircraft-wing model.
The work highlights an important emerging question at the intersection of machine learning and fluid dynamics: when an RL agent learns through a computational environment, how much of what it learns reflects general principles of the underlying physics, and how much is shaped by the environment in which it was trained? Geometry, boundary conditions, accessible flow regimes, numerical resolution, observations, actions, and rewards all influence what an agent can experience and learn.
HydroGym provides a framework for systematically investigating these questions and represents an important step toward more generalizable and intelligent approaches to computational flow control in real-world applications.