Uni-Flow: a unified autoregressive-diffusion model for complex multiscale flows

Xiao Xue; Tianyue Yang; Mingyang Gao; Leyu Pan; Maida Wang; Kewei Zhu; Shuo Wang; Jiuling Li

physics.flu-dynFeb 17, 2026

Uni-Flow: a unified autoregressive-diffusion model for complex multiscale flows

Authors: Xiao Xue, Tianyue Yang, Mingyang Gao, Leyu Pan, Maida Wang, Kewei Zhu, Shuo Wang, Jiuling Li, +2 more

Organizations: Centre for Computational Science, University College London, London, UK · Department of Earth Science and Engineering, Imperial College London, London, UK · Department of Chemical Engineering, University College London, London, UK · Department of Physics, Eindhoven University of Technology, Eindhoven, Netherlands · School of Civil & Environmental Engineering, Queensland University of Technology, Brisbane, Australia · Australian Centre for Water and Environmental Biotechnology, The University of Queensland, Brisbane, Australia · Institute for Mechanics, Computational Mechanics Group, Technical University of Darmstadt, Germany · Centre for Advanced Research Computing, University College London, London, UK

Abstract

Spatiotemporal flows govern diverse phenomena across physics, biology, and engineering, yet modelling their multiscale dynamics remains a central challenge. Despite major advances in physics-informed machine learning, existing approaches struggle to simultaneously maintain long-term temporal evolution and resolve fine-scale structure across chaotic, turbulent, and physiological regimes. Here, we introduce Uni-Flow, a unified autoregressive-diffusion framework that explicitly separates temporal evolution from spatial refinement for modelling complex dynamical systems. The autoregressive component learns low-resolution latent dynamics that preserve large-scale structure and ensure stable long-horizon rollouts, while the diffusion component reconstructs high-resolution physical fields, recovering fine-scale features in a small number of denoising steps. We validate Uni-Flow across canonical benchmarks, including two-dimensional Kolmogorov flow, three-dimensional turbulent channel inflow generation with a quantum-informed autoregressive prior, and patient-specific simulations of aortic coarctation derived from high-fidelity lattice Boltzmann hemodynamic solvers. In the cardiovascular setting, Uni-Flow enables task-level faster than real-time inference of pulsatile hemodynamics, reconstructing high-resolution pressure fields over physiologically relevant time horizons in seconds rather than hours. By transforming high-fidelity hemodynamic simulation from an offline, HPC-bound process into a deployable surrogate, Uni-Flow establishes a pathway to faster-than-real-time modelling of complex multiscale flows, with broad implications for scientific machine learning in flow physics.

Figures & tables

Figure 1: Uni-Flow framework for multiscale spatiotemporal modelling and representative applications. Uni-Flow decouples temporal evolution and spatial refinement through a unified autoregressive–diffusion formulation. High-resolution physical fields are first downsampled to a low-resolution representation, that retains the large-scale degrees of freedom governing long-horizon temporal evolution relevant to the target observables. The AR component may be instantiated using different operator-learning architectures, such as neural operators or Koopman-based models, to ensure stable long-horizon temporal dynamics. At selected timesteps, the low-resolution state is upscaled and provided as conditioning to a diffusion-based refinement model, which reconstructs high-resolution physical fields and recovers fine-scale spatial structure through a small number of denoising steps. The lower panels illustrate representative applications of Uni-Flow across increasing physical and modelling complexity: two-dimensional Kolmogorov flow ( $256\times 256$ ), three-dimensional turbulent channel inflow generation ( $192\times 192$ ), and patient-specific stenotic aortic flow, where surface pressure fields are parameterised on a two-dimensional UV domain ( $512\times 512$ ).

Figure 2: Evaluation of the Uni-Flow model on the two-dimensional Kolmogorov flow. Panel (a) shows sequential vorticity snapshots for the ground truth (top), low-resolution autoregressive (LR-AR) prediction (middle), and absolute error (bottom), where LR-AR captures the dominant shear-layer dynamics but under-resolves fine-scale vortices. Panel (b) presents the time-averaged vorticity field, showing that the LR-AR model reproduces the large-scale roll structures imposed by the Kolmogorov forcing. Panel (c) displays the temporal autocorrelation, indicating stable short-term dynamics with gradual decay at longer horizons. Panel (d) shows the time-averaged kinetic energy spectrum $\langle E(k)\rangle$ , demonstrating consistent inertial scaling but slight underestimation at high wavenumbers. Panel (e) presents the Q-Q comparison between predicted and ground-truth vorticity distributions, confirming statistical alignment with minor deviation in the tails.

Figure 3: Demonstration of Uni-Flow for generating turbulent inflow conditions in a channel-flow configuration. Panel (a) presents the time-averaged velocity comparison between Reference, LR-AR, and Uni-Flow over 600 time frames. Panel (b) shows instantaneous streamwise velocity fields for the reference LBM-LES (left), LR-AR baseline (middle), and Uni-Flow prediction (right). Panel (c) displays the streamwise energy spectrum, confirming that Uni-Flow recovers both the inertial and dissipative ranges with higher fidelity than LR-AR. The blue, red and green lines are representing reference data, Uni-Flow, and LR-AR cases respectfully. Panel (d) shows the time-averaged normalised turbulence velocity profiles, where Uni-Flow matches the reference at all scales, whereas, LR-AR failed to match near the wall. Panel (e) presents the probability density function of normalised velocity distribution, demonstrating that Uni-Flow reproduces the correct velocity distribution whereas LR-AR case failed to capture the low velocity region.

Figure 4: Uni-Flow learning of spatiotemporal pressure dynamics in aortic stenosis. Panel (a) Overview of the Uni-Flow framework for complex cardiovascular flows. Three-dimensional numerical simulations (HemeLB) of a stenotic aorta are mapped to a two-dimensional UV domain, where Uni-Flow learns the spatiotemporal evolution of wall pressure fields entirely in 2D. The predicted fields are then inversely mapped to the three-dimensional geometry to reconstruct the physical distribution on the aortic surface. Panel (b) Comparison of wall pressure evolution across a cardiac cycle at 5 representative time frames ( $t_{1}$ to $t_{5}$ ). Uni-Flow accurately reproduces the high-pressure buildup upstream of the stenosis and the downstream recovery region, closely matching the ground-truth simulation results both spatially and temporally.

Figure 5: Quantitative assessment of pressure dynamics in stenotic aortic flow. Panel (a) shows the ensamble average surface pressure distributions on the aortic wall for the ground truth (left), LR-AR (middle), and Uni-Flow (right), where Uni-Flow accurately captures the high-pressure region upstream of the stenosis and the recovery downstream. Panel (b) presents the temporal autocorrelation functions, showing that Uni-Flow maintains periodic coherence across cardiac cycles. Panel (c) displays the pressure kernel density estimate plots, indicating that Uni-Flow reproduces the sharp, narrow distribution observed in the HemeLB reference. Panel (d) shows the pressure energy spectra for the peak of heart beat cycle, demonstrating that Uni-Flow recovers the correct spectral scaling and dissipation range across all resolved wavenumbers, while LR-AR model failed to capture the high wave number energy spectrum. Panel (e) quantifies the pressure-drop characteristics within the highlighted region over 8 cardiac cycles (200 inference frames), comparing peak amplitudes, statistical distributions, and temporal waveforms. Uni-Flow remains within the simulated physiological range of 90-115 mmHg, accurately reproducing pulsatile amplitude and phase while matching the mean and variability of the reference. In contrast, the LR-AR baseline underestimates systolic peaks and shows reduced dynamic range.

Explore similar work

One Scale at a Time: Scale-Autoregressive Modeling for Fluid Flow Distributions
Apr 13, 2026Mario Lino, Nils ThuereyGenerative Modeling Autoregressive Modeling
Technical University of Munich
Towards Scalable One-Step Generative Modeling for Autoregressive Dynamical System Forecasting
May 7, 2026Tianyue Yang, Xiao XuePDE Surrogate Modeling One-Step Generative Modeling
The Center for Computational Science University College London
Physical Fidelity Reconstruction via Improved Consistency-Distilled Flow Matching for Dynamical Systems
May 7, 2026Sicheng Ma, Tianyue Yang, Xiuzhe Wu +1Flow Matching One-Step Generative Modeling
Department of Physics University of Cambridge · The Center for Computational Science University College London · Department of Computer Science University College London

Apr 13, 2026J/K move · Enter open · S save

One Scale at a Time: Scale-Autoregressive Modeling for Fluid Flow Distributions

Mario Lino, Nils Thuerey

Technical University of Munich

Generative Modeling Autoregressive Modeling Computational Fluid Dynamics

Open paper AskPDF arXiv

Analyzing unsteady fluid flows often requires access to the full distribution of possible temporal states, yet conventional PDE solvers are computationally prohibitive and learned time-stepping surrogates quickly accumulate error over long rollouts. Generative models avoid compounding error by sampling states independently, but diffusion and flow-matching methods, while accurate, are limited by the cost of many evaluations over the entire mesh. We introduce scale-autoregressive modeling (SAR) for sampling flows on unstructured meshes hierarchically from coarse to fine: it first generates a low-resolution field, then refines it by progressively sampling higher resolutions conditioned on coarser predictions. This coarse-to-fine factorization improves efficiency by concentrating computation at coarser scales, where uncertainty is greatest, while requiring fewer steps at finer scales. Across unsteady-flow benchmarks of varying complexity, SAR attains substantially lower distributional error and higher per-sample accuracy than state-of-the-art diffusion models based on multi-scale GNNs, while matching or surpassing a flow-matching Transolver (a linear-time transformer) yet running 2-7x faster than this depending on the task. Overall, SAR provides a practical tool for fast and accurate estimation of statistical flow quantities (e.g., turbulent kinetic energy and two-point correlations) in real-world settings.

Abstract

Explore similar work

One Scale at a Time: Scale-Autoregressive Modeling for Fluid Flow Distributions

Towards Scalable One-Step Generative Modeling for Autoregressive Dynamical System Forecasting

Physical Fidelity Reconstruction via Improved Consistency-Distilled Flow Matching for Dynamical Systems