Rectified Flows for Fast Multiscale Fluid Flow Modeling

Victor Armegioiu, Yannick Ramic, Siddhartha MishraSubmission to ICLR 2026 · Rejected

TL;DR

Simulating chaotic, multiscale fluid flows with generative diffusion models yields high physical fidelity but requires hundreds of slow denoising steps. This paper introduces RecFlow, a framework adapting rectified flows to conditional PDEs that learns straight ODE trajectories from noise to solution states. Paired with a curvature-aware adaptive ODE integrator, RecFlow matches diffusion-quality statistics and energy spectra with just 8–10 solver steps—delivering up to a 22× inference speedup and converging significantly faster during training.

What Problem Does This Paper Address?

Simulating multiscale, chaotic fluid flows (like turbulent mixing or shock-wave interactions) requires resolving tiny vortices and high-frequency wave interactions. Standard deterministic surrogate models (like Fourier Neural Operators, FNO) trained via Mean Squared Error (MSE) suffer from "spectral collapse," over-smoothing fine-scale turbulence into an unphysical blur.

Recent generative surrogates based on conditional score-based diffusion (e.g., GenCFD) fix this by learning posterior distributions over flow states. However, diffusion models require simulating reverse-time stochastic differential equations (SDEs) over 100+ discretization steps. Because each step evaluates an expensive high-resolution neural network, real-time prediction and large-ensemble uncertainty quantification (UQ) remain computationally prohibitive.

Methods

The authors replace the curved stochastic reverse-SDE path of diffusion models with a deterministic Rectified Flow (RecFlow):

  • Straight Transport via Rectified Flows: Instead of wandering through noise space, RecFlow learns an ODE velocity field $v_\theta(u_\tau, u_i, \tau)$ that pushes probability distributions along nearly straight linear interpolations $X_\tau = (1-\tau)u_0 + \tau u_1$. Straighter paths mean numerical ODE solvers can take drastically larger steps without accumulating truncation errors.
  • Curvature-Aware Integration: The authors introduce an adaptive solver mechanism requiring no extra model training:
  • An Exponential Moving Average (EMA) tracks velocity trends: $v^{\text{ema}}_t = \lambda v^{\text{ema}}_{t-\Delta \tau} + (1-\lambda) v_t$.
  • A straightness score $s_t = \|v_t - v^{\text{ema}}_t\|_2$ flags regions where the trajectory curves.
  • When $s_t$ spikes, a Tikhonov-regularized convex blend gently smooths the predicted velocity towards $v^{\text{ema}}$, and an adaptive step-size rule contracts the step size $\Delta \tau_t \propto 1 / \sqrt{\kappa_1 s_t + \kappa_2}$ to keep local truncation error bounded.
  • Architecture: A UViT backbone (a U-Net with multi-head self-attention at the bottleneck and Feature-wise Linear Modulation / FiLM temporal conditioning) predicting the displacement field.

Datasets

The authors benchmark across three 2D multiscale fluid mechanics problems:

  1. SL2D (Shear Layer): Incompressible 2D Navier–Stokes system modeling shear-layer roll-up into vortices (spectral viscosity solver at $512^2$, downsampled to $128^2$; 79,200 samples).
  2. CS2D (Cloud-Shock Interaction): Compressible Euler equations modeling a planar shock wave striking a dense cloud (finite-volume ALSVINN solver at $256^2$, downsampled to $128^2$; 40,000 samples).
  3. RM2D (Richtmyer–Meshkov): Compressible Euler equations simulating shock-driven fluid interface instability and turbulent mixing (ALSVINN solver at $256^2$, downsampled to $128^2$; 64,000 samples).

Evaluation uses an ensemble-perturbation protocol: generating micro-ensembles (1,000 perturbed initial conditions within an $\varepsilon$-ball) evolved via high-fidelity PDE solvers to construct empirical ground-truth posterior distributions for chaotic regimes.

Key Insights and Ideas

  • Straightening beats diffusion ODE tricks: Running a probability-flow ODE (PF-ODE) or standard non-rectified flow matching on pre-trained diffusion networks does not solve the problem—their velocity fields remain stiff and still require large step counts ($>8$) to capture spectra accurately. Rectifying the transport trajectories during training is what genuinely linearizes the dynamics.
  • Curvature as a controller, not a detector: While prior diffusion work uses trajectory curvature for out-of-distribution (OOD) detection, RecFlow uses curvature as an active runtime control knob to regularize updates and adapt step sizes dynamically.
  • Training efficiency gains: Rectified flows converge much faster than diffusion models in compute time. RecFlow outperformed a capacity-matched (22M parameter) GenCFD model across all metrics by ~100k training iterations, while GenCFD required 300k–600k iterations to reach comparable accuracy on several metrics (and never matched RecFlow on cross-velocity Wasserstein distance).

Key Results

  • Inference Speed: RecFlow requires only 8–10 ODE integration steps (vs. $\ge 128$ steps for GenCFD / diffusion baselines), achieving up to a 22× wall-clock speedup.
  • Statistical and Spectral Fidelity:
  • On CS2D, RecFlow achieved a 1-point Wasserstein distance ($W_1$) of $0.0091$ on density $\rho$ (compared to GenCFD's $0.0151$ and FNO's $0.1455$).
  • On SL2D, RecFlow reached $W_1 = 0.0214$ on $u_x$ and $0.0164$ on $u_y$, outperforming both GenCFD ($0.0276, 0.0213$) and deterministic FNO ($0.184, 0.150$).
  • RecFlow captures high-wavenumber energy spectra in as few as 5 steps, accurately reproducing power-law turbulent decay where score-based diffusion remains noisy and under-resolved.
  • Curvature-Aware Integration Impact: Applying the curvature-aware integration scheme produced an additional 2.6% error reduction on CS2D and 6.0% error reduction on SL2D over standard Euler integration.

Limitations and Future Work

  • 2D Benchmarks: Validation is currently restricted to 2D incompressible and compressible flows; extending the framework to full 3D turbulence and multiphysics systems remains future work.
  • Discretization and Architecture: While UViT scales well, developing more lightweight architectures for ultra-fast, real-time edge deployment is still an open area.
  • UQ Integration: Further work is needed to integrate RecFlow directly into active uncertainty-quantification and PDE-constrained optimization pipelines.

Summary updated

Abstract

The statistical modeling of fluid flows is very challenging due to their multiscale dynamics and extreme sensitivity to initial conditions. While recently proposed conditional diffusion models achieve high fidelity, they typically require hundreds of stochastic sampling steps at inference. We introduce a rectified-flow framework that learns a time-dependent velocity field, transporting input to output distributions along nearly straight trajectories. By casting sampling as solving an ordinary differential equation (ODE) along this straighter flow field, our method makes each integration step much more effective, using as few as eight steps versus (more than) 128 steps in standard score-based diffusion, without sacrificing predictive fidelity. In addition, we develop a curvature-aware integration scheme that monitors local path straightness and adaptively regularizes the velocity and step size, improving stability and accuracy at essentially no training cost. Experiments on challenging multiscale flow benchmarks show that rectified flows recover the same posterior distributions as diffusion models, preserve fine-scale features that MSE-trained baselines miss, and deliver high-resolution samples in a fraction of inference time.
BibTeX
@misc{armegioiu2025rectified,
  author = {Victor Armegioiu and Yannick Ramic and Siddhartha Mishra},
  title = {{Rectified Flows for Fast Multiscale Fluid Flow Modeling}},
  year = {2025},
  eprint = {2506.03111},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.03111},
}

Loading related content...

Figure