Model predictive control (MPC), also called receding horizon control, works like this: at every step the controller simulates the future over a fixed window with a learned or physical model, picks the best action sequence, applies only the first action, then repeats. Flow-control papers that use it pick the window length by hand. The reason this matters for fluids is the Reynolds number ($Re$), the ratio of inertial to viscous forces. As $Re$ grows, flows break into finer, faster, more chaotic structures. One might therefore expect the useful look-ahead to shrink, so that a horizon tuned on a cheap low-$Re$ simulation would be too long at high $Re$.
The opposite is also plausible. Evidence from two-dimensional turbulence suggests that how far ahead a flow can be predicted is set by large coherent vortices, not by small eddies. If that holds, the right horizon measured in physical time should stay roughly constant across $Re$. We propose to measure which of these is true for learned-model MPC on standard flow-control benchmarks. We then build and test a cheap horizon rule that needs no reinforcement learning: the controller sets its own look-ahead from how fast its ensemble of world models starts to disagree. Because this rule is calibrated on the model's own uncertainty, it should transfer to Reynolds numbers too expensive to tune by brute force.
Core Claim
The closed-loop optimal MPC horizon $N^*(Re)$, expressed in convective time units (body diameter divided by free-stream speed), follows the lifetime of the large coherent structures being controlled. It does not follow the small-scale chaotic time scale, which shrinks like $Re^{-1/2}$ for the Kolmogorov time. A horizon rule that ends the look-ahead when model predictions stop being trustworthy recovers most of the benefit of tuning the horizon separately at every $Re$, at none of the tuning cost.
Concretely, let $f_1,\dots,f_M$ be an ensemble of learned dynamics models, let $x$ be the current estimated flow state under a candidate action sequence, and let $D(h\mid x)$ be the normalized spread of their $h$-step predictions of the controlled quantity, for example drag or lift. The rule is
$$H(x) = \min\{\, h \le N_{\max} : D(h \mid x) > \varepsilon \,\},$$
where $\varepsilon$ is a single threshold calibrated once at the lowest training $Re$. The scaling question is whether $N^*(Re)\,U/L$ is flat or decays as $Re$ rises. Here $U$ is the free-stream or bulk velocity and $L$ is the body or channel length.
Why This and Not the Obvious Combination
The obvious idea at this intersection is to port Reinforcement Learning of the Prediction Horizon in Model Predictive Control to fluids. That paper trains a Soft Actor-Critic agent to pick the horizon online. On an inverted pendulum and a unicycle obstacle task, it beat the best-tuned fixed horizon by 4% and 8%. Carrying it over to flows would add a second learning loop on top of an expensive simulator, and it would not explain why a given horizon is right. Our proposal differs in two ways. First, it treats the horizon as a physical measurement to be explained by flow time scales. Second, it replaces the learned horizon policy with an uncertainty rule that has one scalar parameter. That rule is the thing we expect to extrapolate.
Prior fluid MPC fixes the window. Deep Dynamical Modeling and Control of Unsteady Fluid Flows learns a linear latent (Koopman) model of a cylinder wake. It keeps prediction error under 0.2% over 128 steps and suppresses shedding with linear MPC, but only at $Re=50$. It is our method source for recursively trained latent models and our motivation, since its horizon was never revisited at higher $Re$. Physics-informed Neural-operator Predictive Control predicts channel-flow evolution with a physics-regularized neural operator and extrapolates to unseen $Re_b=15{,}000$, but it does so with a fixed predictive structure. The physical hypothesis comes from Yuxuan Yang et al.. They found that in 2D Kolmogorov flow at $Re=100$, extreme-event predictability ranges from 1 to over 4 Lyapunov times, is tied to the lifetime of quadrupolar precursor vortices, and is insensitive to scales below $\lambda_c\approx 0.3L$. The opposing view comes from the scaling literature: Sakata et al., Eshbal et al., and Lee document separation shifts, wake fragmentation, and changed shedding behavior as $Re$ rises. Our experiment decides between these two views in a control setting. A further constraint is that the Yang et al. result comes from 2D flow, which lacks vortex stretching. We therefore include a 3D configuration.
Research Program
We work in HydroGym. Its JAX-based differentiable backend lets us solve the MPC problem by gradient-based shooting through a learned model. We use three families of environments:
- Cylinder. The 2D cylinder from $Re=100$ upward, and the 3D cylinder over $Re\in\{100,200,300,400\}$, the range used by Suárez et al..
- Kolmogorov flow. 2D Kolmogorov flow swept from about $Re=40$ to several hundred, which connects to the Yang et al. predictability analysis.
- Channel. The 3D turbulent channel, as the wall-bounded case.
World model. In each environment we train an ensemble of five recursively trained latent dynamics models on data from the lower part of the $Re$ range, with $Re$ given as an input. We follow the multi-step training of Morton et al. In the channel we also try a neural-operator variant following PINO-PC.
Step 1: map the horizon-versus-performance curve. We sweep fixed horizons at every $Re$, including held-out $Re$ above the training range. This yields $N^*(Re)$ and a performance-versus-horizon curve with its compute cost.
Step 2: test candidate time scales against $N^*$. We compare the measured optimal horizons with:
- the natural shedding period, or the eddy turnover time in the channel;
- the model ensemble's error-doubling time;
- the leading Lyapunov time, estimated from twin simulations;
- the Kolmogorov time.
Step 3: run the uncertainty rule $H(x)$. We calibrate $\varepsilon$ at the lowest $Re$, freeze it, and evaluate across the full range.
Step 4: low-pass filtering test. This addresses the tension between coherent structures and small-eddy chaos directly. We restrict the MPC cost and the model state to scales above $\lambda_c$, and check whether control quality and the horizon trend survive as $Re$ increases.
Constraints throughout. All controllers enforce actuator magnitude and rate limits, and zero net mass flux, as hard constraints in the optimization. Without this, horizon effects would be confounded with how each method handles constraints.
Baselines
The naive reference is no actuation. All drag and lift-fluctuation numbers are reported relative to the uncontrolled flow at the same $Re$. Alongside it we run opposition control in the channel, and the best open-loop periodic forcing in the cylinder cases, with forcing frequency swept per $Re$. Both are cheap and familiar, and a predictive method must clearly beat them to justify itself. Discovering drag reduction strategies in wall-bounded turbulent flows using deep reinforcement learning reports 26% drag reduction for opposition control versus 43% for learned control in the minimal channel at $Re_\tau=180$.
The strongest model-free competitor is PPO trained at the low-$Re$ end with multi-agent weight sharing, following Suárez et al.. They report 17.2% drag reduction at $Re=200$ and 15.1% at $Re=400$ on the 3D cylinder. We run PPO both zero-shot at held-out $Re$ and with a fine-tuning budget equal to the simulator time our method spends on calibration.
The strongest predictive competitor in the channel is PINO-PC. It reports 39.0% drag reduction at unseen $Re_b=15{,}000$, 9.0 points above DDPG and 26.5 points above MP-CNN. Where our channel setup can match its configuration, we compare against that number directly. Otherwise we rerun its fixed-horizon structure inside our pipeline, using the same model.
The horizon baselines isolate the key ingredient. They all share the same world-model ensemble, solver, actuator constraints, and per-step compute cap:
- a fixed horizon tuned at the lowest $Re$ and reused unchanged;
- a fixed horizon rescaled by the convective time ratio;
- the RL horizon selector of Bøhn and Gros, trained with the same simulator budget;
- an oracle that grid-searches the best fixed horizon separately at each test $Re$.
Beating the baselines means two things. Without any per-$Re$ tuning, $H(x)$ must close at least two thirds of the gap between the reused low-$Re$ horizon and the oracle at held-out $Re$. It must also match or beat the RL selector at lower compute. Separately, MPC with $H(x)$ must exceed zero-shot PPO in drag reduction at the highest held-out $Re$.
Why not just…?
Why not just rescale the horizon by the shedding period or convective time? This is the most serious threat. If the coherent-structure hypothesis is right, a physics-based rescaling may be all that is needed. The rescaled fixed horizon is therefore an explicit baseline. If it matches $H(x)$, we report that as a positive finding about scaling, not as a failure. The uncertainty rule adds value only where no single dominant period exists, such as the channel or 2D Kolmogorov flow, or where the model's own reliability drifts with $Re$. The experiment shows whether that happens.
Why not just learn the horizon with reinforcement learning, as Bøhn and Gros did? Their 4–8% gains came on low-dimensional systems with fast simulation. In flows, every training episode is a costly CFD run, and a learned horizon policy trained at low $Re$ faces the same transfer problem that Guastoni et al. and PINO-PC document for model-free controllers. We run it under an equal simulator budget, so the comparison is direct.
Why not just use the longest horizon compute allows? Morton et al. show that poorly trained models diverge after about 32 steps. Long horizons then optimize against fiction, and the cost grows linearly with the horizon. Step 1 measures directly whether performance plateaus or drops beyond $N^*$.
Why not skip MPC and use model-free reinforcement learning? PINO-PC's 9-point margin over DDPG at unseen $Re$ suggests predictive control generalizes better. The Bellman-Jensen gap analysis of Lu et al. gives a reason to expect look-ahead to help, though it was shown in energy dispatch, not flows. We test this on cylinder flows with matched budgets, a setting where no reported comparison was found on Lacuna.
Proposed Experiments
Experiment 1: the scaling test. We plot $N^*(Re)\,U/L$ for the cylinder and Kolmogorov sweeps, with a bootstrap confidence band. The hypothesis fails if the slope versus $\log Re$ is near $-1/2$, the Kolmogorov-time scaling. It is supported if the slope is near zero and tracks the shedding or precursor-structure lifetime. Repeating the test on the 3D cylinder checks whether vortex stretching changes the answer.
Experiment 2: extrapolation. We calibrate on the lower half of each $Re$ range and test above it. We report drag reduction, lift RMS, constraint violations, and wall-clock time per control step. The measure of interest is regret relative to the oracle horizon, compared across $H(x)$, rescaled-fixed, and the RL horizon.
Experiment 3: the filtering test. We run MPC with the cost and model restricted to scales above $\lambda_c$. If control quality degrades with $Re$ while it holds in the unfiltered version, small-scale chaos matters for control even when it does not matter for predictability. That would directly contradict extending the Yang et al. finding to control.
Experiment 4: scope check. A negative result would come from the channel. If $H(x)$ fails there while succeeding on bluff bodies, the method's scope is limited to flows with a dominant coherent time scale, and we would say so.
References
- Reinforcement Learning of the Prediction Horizon in Model Predictive Control
- Deep Dynamical Modeling and Control of Unsteady Fluid Flows
- Physics-informed Neural-operator Predictive Control for Drag Reduction in Turbulent Flows
- Hierarchy of extreme-event predictability in turbulence revealed by machine learning
- A Surface-Mounted Circular Cylinder in a Turbulent Boundary Layer : Flow Visualization and Surface-Pressure Properties
- Three-dimensional flow field measurements in the wake of a tethered sphere crossing the onset of vortex induced vibrations
- Free end effects on the near wake flow structure behind a finite circular cylinder
- HydroGym: A Reinforcement Learning Platform for Fluid Dynamics
- Active flow control for three-dimensional cylinders through deep reinforcement learning
- Discovering drag reduction strategies in wall-bounded turbulent flows using deep reinforcement learning
- Reinforcement Learning with Imperfect Transition Predictions: A Bellman-Jensen Approach