Physics Foundation Models for Efficient Control of Fluid Flows

Mahindra Rautela; Shaifalee Saxena; Alexander Scheinker

physics.flu-dynOct 10, 2026

Physics Foundation Models for Efficient Control of Fluid Flows

Authors: Mahindra Rautela, Shaifalee Saxena, Alexander Scheinker

Organizations: Los Alamos National Laboratory, Los Alamos, NM, USA · University of New Mexico, Albuquerque, NM, USA

Abstract

We investigate whether a physics foundation model (PFM) pretrained for PDE forecasting can be repurposed for closed-loop fluid-flow control. We construct a PFM-based controller by attaching lightweight Gaussian actor and critic heads to frozen pretrained representations and train the policy using proximal policy optimization (PPO). We evaluate the approach on HydroGym's two-dimensional Kolmogorov-flow environment using vorticity observations at $64\times64$ and $128\times128$ resolutions and Reynolds numbers $\mathrm{Re}=100$ and $200$ . Across these settings, we compare the frozen PFM controller against a baseline MLP policy trained with the same PPO framework. The PFM controller achieves consistently higher reward while requiring smaller executed action magnitudes over training, indicating improved control performance with reduced control effort. These results provide initial evidence that representations learned through self-supervised next-frame prediction can be transferred directly to closed-loop fluid flow control. Project page: https://mahindrautela.github.io/morph.

Figures & tables

Figure 1: Repurposing a physics foundation model for closed-loop flow control. The latent state representation is passed to lightweight actor and critic heads, which parameterize a Gaussian control policy and estimate the state value, respectively. PPO updates the control heads from closed-loop interaction with the Kolmogorov-flow environment, while the pretrained foundation-model backbone remains frozen.

Environment	Value	PPO	Value
Reynolds number	$\{100,200\}$	Baseline / MORPH head LR	$10^{-4}$
Grid	$64\times 64,\;128\times 128$	Parallel envs	8
Domain	$[0,2\pi]^{2}$	Rollout steps / update	10
$\Delta t$	$10^{-3}$	Minibatches	8
Control interval	10.0	Update epochs	10
Action modes	$4,5,6,7$	Discount $\gamma$	0.99

Table 1: Kolmogorov-flow and PPO configuration used for the reported comparisons. The same control interval, reward, and PPO optimization settings are used across resolutions and Reynolds numbers.

Panel of Figure 2 — Figure 2: Combined PPO training results across Reynolds number and spatial resolution. Top: $\mathrm{Re}=100$ ; bottom: $\mathrm{Re}=200$ . Left: reward versus PPO update. Right: mean executed action magnitude, computed as the mean $\ell_{1}$ norm of the clipped four-dimensional control action. At $\mathrm{Re}=100$ , baseline and frozen-MORPH results are shown for both $64\times 64$ and $128\times 128$ grids. Reward at each PPO update is the sum over its 10 rollout steps, with rewards averaged across the eight parallel environments at each step. All curves are smoothed using a 50-update moving average for visualization.

Panel of Figure 3 — Figure 3: Closed-loop rollouts of the trained frozen MORPH-Ti controller. The snapshots show real-space vorticity at $t=0,50,100,150,200$ . With a control interval of 10 simulation-time units, each trajectory spans 20 policy decisions.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Model	Pretraining input	Scale	History	KF in pretraining	Control adaptation
MORPH-Ti	$s_{t}$	7M	No	No	Direct
Poseidon-T	$(s_{t},\Delta t)$	20.8M	No	No	Input/time adaptation
DPOT-Ti	$s_{t-9:t}$	7M	Yes	No	History required
Walrus	$s_{t-5:t}$	1.3B	Yes	No	History required
PDE-Transformer-S	$s_{t}$	33.2M	No	Yes	Direct

Appendix

Table 3: Comparison of representative PDE foundation models for the present single-state control setting. The comparison concerns compatibility with state-to-action adaptation rather than forecasting performance. MORPH provides a compact backbone whose pretrained input and temporal interfaces can be retained without introducing an explicit history or lead-time variable.

Physical Foundation Models

See all 14 papers

The HydroGym Reinforcement Learning Platform for Fluid Dynamics
Dec 19, 2025Christian Lagemann, Sajeda Mokbel, Miro Gondrum +18RL Control RL Benchmarks
Department of Mechanical Engineering, University of Washington, Seattle, WA 98195, United States · AI Institute in Dynamic Systems, University of Washington, Seattle, WA 98195, United States · Chair of Fluid Mechanics and Institute of Aerodynamics, RWTH Aachen University, Aachen, Germany +9
Physics and Data Driven Transformer-Mamba Framework for Flow Field
Sep 24, 2026Zhuo Zhang, Shun Zou, Canqun Yang +1PDE Surrogate Modeling PDE Operator Learning
College of Computer Science and Technology, National University of Defense Technology, Changsha, China · College of Artificial Intelligence, Nanjing Agricultural University, NanJing, China · National SuperComputer Center in Tianjin, Binhai New Area, Tianjin, China +1
Heuristic Learning for Active Flow Control Using Coding Agents
Jul 13, 2026Paul Garnier, Jonathan Viquerat, Elie HachemRL Control Language Model-Based Control
Mines Paris - PSL University · Centre for Material Forming (CEMEF) · CNRS +1

Dec 19, 2025J/K move · Enter open · S save

The HydroGym Reinforcement Learning Platform for Fluid Dynamics

Christian Lagemann, Sajeda Mokbel, Miro Gondrum, Mario Rüttgers+17

Department of Mechanical Engineering, University of Washington, Seattle, WA 98195, United States · AI Institute in Dynamic Systems, University of Washington, Seattle, WA 98195, United States · Chair of Fluid Mechanics and Institute of Aerodynamics, RWTH Aachen University, Aachen, Germany+9

RL Control RL Benchmarks Reinforcement Learning Computational Fluid Dynamics

Open paper AskPDF arXiv

Modeling and controlling fluids is critical across science and engineering. Effective flow control can increase lift, reduce drag, enhance mixing, and attenuate noise, potentially unlocking new technologies. Yet controlling fluids is hard: the dynamics are high-dimensional, nonlinear, and multiscale. While reinforcement learning (RL) has recently succeeded in robotics and protein folding through shared benchmarks, fluid dynamics has resisted such progress: each controller is typically tuned to a single geometry and operating point, making results hard to accumulate, transfer, and compare. We introduce HydroGym, a solver-independent RL platform for flow control, and show that standardized infrastructure unlocks transferable control intelligence across flow regimes. HydroGym provides 61+ validated environments spanning laminar to turbulent flows, with systematic Reynolds number progressions up to Re=400,000 and Mach number variations in 2D and 3D. It supports diverse backends, including finite-volume, spectral-element, finite-element, lattice-Boltzmann, and fully differentiable solvers for gradient-enhanced optimization. Across environments, RL agents consistently discover robust control principles, such as boundary-layer manipulation, acoustic-feedback disruption, and wake reorganization, yielding drag reductions exceeding 90% in canonical configurations. Critically, we demonstrate zero-shot transfer: agents trained only on a simplified channel flow achieve 38% friction-drag reduction on an unseen 3D wing section at chord Reynolds number Re=200,000 reducing exploration costs by four orders of magnitude versus direct on-wing optimization. This suggests RL agents uncover essential physics rather than configuration-specific patterns, pointing toward generalizable control. HydroGym offers extensible, scalable community infrastructure for fluid dynamics, machine learning, and control research.

Abstract

Physical Foundation Models

Explore similar work

The HydroGym Reinforcement Learning Platform for Fluid Dynamics

Physics and Data Driven Transformer-Mamba Framework for Flow Field

Heuristic Learning for Active Flow Control Using Coding Agents