Controlling complex fluid flows with reinforcement learning and differentiable solvers

The control of fluid flows is central to many areas of science and engineering, including energy, transportation, and medicine. Effective flow control can, for instance, increase aerodynamic lift, reduce drag in automotive systems, and improve mixing in combustion. However, fluid systems are inherently high-dimensional, nonlinear, and multiscale, making traditional control methods computationally challenging. Reinforcement learning (RL) has shown transformative potential across several domains, from robotics to molecular design, but its progress in fluid dynamics has been limited by a lack of standardized benchmarks, as well as accessibility and computational barriers. In this talk, I will present our work on HydroGym, a RL platform for fluid dynamics. HydroGym provides reproducible environments for studying fluid control for several 2D and 3D systems with applications in engineering. By incorporating differentiable solvers into the learning loop, we further demonstrate that RL can exploit environment gradient information to achieve improved sample efficiency and discover high-performing control policies. This talk highlights how advances in reinforcement learning are shaping the future of flow control in computational fluid dynamics, and, more broadly, how differentiable solvers are redefining the optimization process for complex fluid systems.