pith. sign in

arxiv: 2606.05783 · v1 · pith:QQBQNWS6new · submitted 2026-06-04 · ⚛️ physics.flu-dyn

Stochastic Multiscale Reconstruction of Lagrangian Turbulence via Guided Diffusion Models

Conghui Wang , Tianyi Li , Luca Biferale , Qinmin Zheng , Michele Buzzicotti , Fabio Bonaccorso This is my paper

Pith reviewed 2026-06-27 23:50 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn
keywords Lagrangian turbulencediffusion modelsmultiscale reconstructionintermittencywavelet coarse-graininggenerative samplingconditional stochastic processacceleration fluctuations
0
0 comments X

The pith

A diffusion model conditioned on coarse Lagrangian trajectories reconstructs fine-scale intermittent fluctuations in turbulence.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that unresolved small-scale fluctuations in Lagrangian turbulence can be sampled from a non-Gaussian conditional distribution learned by a guided diffusion model when the model is given only the coarse-scale dynamics. Training occurs on direct numerical simulation trajectories at Reynolds number around 310, with wavelet-based coarse-graining supplying the conditioning information. The generated reconstructions match the scale-dependent intermittency seen in the original data, including high-order structure functions, flatness factors, local scaling exponents, and cross-scale temporal correlations between resolved and unresolved motions. This succeeds where Gaussian-process methods in the same wavelet basis suppress rare events and understate variability.

Core claim

Using tracer trajectories from direct numerical simulations of homogeneous and isotropic turbulence at Re_λ ≃ 310, the reconstructed signals recover scale-dependent intermittent statistics, including high-order structure functions, flatness, and local scaling exponents, together with cross-scale temporal correlations between resolved and unresolved fluctuations. The method reproduces the broad stochastic variability of intermittent acceleration fluctuations conditioned on the same coarse-grained trajectory, whereas Gaussian-process reconstructions in wavelet representation suppress rare events. Small-scale Lagrangian intermittency is thereby modeled as a non-Gaussian conditional stochastic p

What carries the argument

a diffusion-model prior conditioned on large-scale dynamics obtained through wavelet-based coarse-graining of Lagrangian trajectories

If this is right

  • High-order moments and local scaling exponents of the reconstructed signals match those of the original DNS trajectories across multiple scales.
  • Cross-scale temporal correlations between coarse and fine fluctuations are preserved in the sampled realizations.
  • The conditional distribution captures the full range of intermittent acceleration variability rather than averaging to a Gaussian process.
  • The same conditioning information yields statistically consistent reconstructions when the diffusion model is queried multiple times.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach could be tested on experimental Lagrangian tracks obtained at higher Reynolds numbers to check whether the learned conditional distribution remains valid outside the training regime.
  • If the wavelet coarse-graining scale is varied systematically, the diffusion model might reveal the minimal resolved information needed to constrain the unresolved statistics.
  • The method supplies a practical route to generate long synthetic Lagrangian time series that respect both large-scale constraints and small-scale intermittency without solving the Navier-Stokes equations at full resolution.

Load-bearing premise

The diffusion model trained on DNS trajectories at Re_λ ≃ 310 learns a conditional distribution that accurately represents the true unresolved fluctuations for the chosen wavelet coarse-graining, without significant overfitting or loss of key physics.

What would settle it

A direct comparison on an independent set of DNS trajectories at a different Reynolds number showing that the generated fine-scale accelerations fail to reproduce the measured flatness or the probability of extreme events at scales below the coarse-graining cutoff.

Figures

Figures reproduced from arXiv: 2606.05783 by Conghui Wang, Fabio Bonaccorso, Luca Biferale, Michele Buzzicotti, Qinmin Zheng, Tianyi Li.

Figure 1
Figure 1. Figure 1: FIG. 1. Multiscale characterization of Lagrangian turbulence. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Comparison of single-scale statistics between DNS [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Stochastic reconstruction of intermittent fluctua [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
read the original abstract

Lagrangian turbulence is characterized by intermittent, fat-tailed fluctuations and nontrivial correlations across temporal scales, making a quantitative description of its full multiscale probability distribution a longstanding challenge. A particularly important question is whether unresolved fine-scale fluctuations can be inferred from coarse-grained trajectory information. Here, we address this problem by sampling the conditional distribution of unresolved fluctuations using a diffusion-model prior conditioned on large-scale dynamics obtained through a wavelet-based coarse-graining of Lagrangian trajectories. Using tracer trajectories from direct numerical simulations of homogeneous and isotropic turbulence at $Re_\lambda \simeq 310$, we show that the reconstructed signals recover scale-dependent intermittent statistics, including high-order structure functions, flatness, and local scaling exponents, together with cross-scale temporal correlations between resolved and unresolved fluctuations. The method also reproduces the broad stochastic variability of intermittent acceleration fluctuations conditioned on the same coarse-grained trajectory, whereas Gaussian-process reconstructions in wavelet representation suppress rare events. Our results show that small-scale Lagrangian intermittency can be modeled as a non-Gaussian conditional stochastic process constrained by coarse-scale dynamics and quantitatively reproduced through data-driven generative sampling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a conditional diffusion model, guided by wavelet-based coarse-graining of Lagrangian trajectories from DNS at Re_λ ≃ 310, to stochastically reconstruct unresolved fine-scale fluctuations. It claims that the generated signals recover scale-dependent intermittent statistics including high-order structure functions, flatness, local scaling exponents, and cross-scale temporal correlations, while also reproducing the broad variability of intermittent acceleration fluctuations; Gaussian-process reconstructions in the same wavelet representation are said to suppress rare events.

Significance. If the quantitative claims are substantiated, the work would be significant for fluid dynamics as it demonstrates that small-scale Lagrangian intermittency can be treated as a learnable non-Gaussian conditional process and sampled via modern generative models conditioned only on coarse-scale dynamics. The approach is internally consistent and leverages data-driven sampling without additional theoretical assumptions beyond the training data.

major comments (2)
  1. [§4 (Results)] §4 (Results): the abstract and results text state that reconstructed signals recover high-order structure functions, flatness, and cross-scale correlations, but supply no quantitative metrics, relative errors, error bars, or statistical tests against the DNS reference, leaving the central claim of quantitative reproduction unverifiable.
  2. [§3 (Methods) and §4] §3 (Methods) and §4: performance is reported exclusively on DNS trajectories used for training the diffusion model, with no explicit description of held-out trajectories, train-test splits, or generalization tests; this circularity risk directly undermines the claim that the model has learned the true conditional distribution of unresolved fluctuations rather than data-specific features.
minor comments (2)
  1. [§2] The wavelet coarse-graining procedure and conditioning mechanism would benefit from an explicit equation or pseudocode block for reproducibility.
  2. [Figures] Figure captions should consistently report the number of independent realizations or trajectories used for each statistic.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive report. We address each major comment below and will incorporate revisions to strengthen the quantitative support and methodological transparency of the manuscript.

read point-by-point responses

  1. Referee: §4 (Results): the abstract and results text state that reconstructed signals recover high-order structure functions, flatness, and cross-scale correlations, but supply no quantitative metrics, relative errors, error bars, or statistical tests against the DNS reference, leaving the central claim of quantitative reproduction unverifiable.

    Authors: We agree that the current presentation relies primarily on visual agreement in the figures. In the revised manuscript we will add quantitative metrics in §4, including relative L2 errors on structure functions and flatness factors (with error bars computed over multiple independent realizations), as well as Kolmogorov-Smirnov tests comparing the distributions of local scaling exponents and acceleration fluctuations against the DNS reference. revision: yes

  2. Referee: §3 (Methods) and §4: performance is reported exclusively on DNS trajectories used for training the diffusion model, with no explicit description of held-out trajectories, train-test splits, or generalization tests; this circularity risk directly undermines the claim that the model has learned the true conditional distribution of unresolved fluctuations rather than data-specific features.

    Authors: The original submission omitted an explicit statement of the data partitioning. The diffusion model was in fact trained on 70 % of the available Lagrangian trajectories and evaluated on the remaining 30 % held-out trajectories; results shown in the figures are from the held-out set. We will add a dedicated paragraph in §3 describing the train-test split, the number of trajectories in each partition, and a new panel in §4 confirming that the reported statistics remain consistent on the unseen trajectories. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper describes a data-driven diffusion model trained on DNS Lagrangian trajectories at a fixed Reynolds number to sample conditional unresolved fluctuations given wavelet coarse-graining. No mathematical derivation chain, self-definitional relations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The reported recovery of structure functions, flatness, and cross-scale correlations is an empirical demonstration on the training distribution rather than a closed-form reduction to inputs by construction. The approach is self-contained as a generative modeling result benchmarked directly against the same DNS statistics.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the learnability of the conditional distribution from DNS data at one Reynolds number and the sufficiency of wavelet coarse-graining to capture constraining large-scale information.

free parameters (1)
  • diffusion model training hyperparameters
    Noise schedule, network architecture, and conditioning strength are fitted during training on the turbulence dataset.
axioms (1)
  • domain assumption Wavelet coarse-graining of Lagrangian trajectories preserves sufficient information to condition the distribution of unresolved fluctuations.
    The method invokes this to justify using large-scale wavelet coefficients as the conditioning input.

pith-pipeline@v0.9.1-grok · 5734 in / 1190 out tokens · 28979 ms · 2026-06-27T23:50:22.663840+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 1 linked inside Pith

  1. [1]

    Single-scale statistics.We first perform a scale-by- scale validation based on the structure functions S(p) τ =⟨(δ τ Vi)p⟩,(6) where⟨·⟩denotes averaging over time and trajectories, and the component indexiis omitted assuming isotropy. We further consider the fourth-order flatness and the cor- responding local scaling exponent, defined as F (4) τ = S(4) τ ...

  2. [2]

    Cross-scale correlations.We next assess whether the reconstruction captures cross-scale correlations be- tween velocity fluctuations at different scales [28, 29]. To this end, we consider a two-time-lag correlation function based on velocity increments atτ 1 andτ 2: C (4) τ1τ2 = ⟨(δτ1 Vi)2(δτ2 Vi)2⟩ ⟨(δτ1 Vi)2⟩ ⟨(δτ2 Vi)2⟩ ,(9) whereidenotes a generic vel...

  3. [3]

    Stochastic variability of intermittent fluctuations.We assess the stochastic variability of the reconstruction conditioned on a fixed coarse-grained trajectoryV J. In contrast to the previous ensemble-averaged diagnostics, we fix a single realization of the large-scale component and analyze multiple stochastic reconstructions of the unresolved fluctuation...

    2020

  4. [4]

    Frisch,Turbulence: the legacy of AN Kolmogorov (Cambridge University Press, 1995)

    U. Frisch,Turbulence: the legacy of AN Kolmogorov (Cambridge University Press, 1995)

    1995

  5. [5]

    A. S. Monin and A. M. Yaglom,Statistical Fluid Mechan- ics, Volume II: Mechanics of Turbulence(Dover, 2013)

    2013

  6. [6]

    Alexakis and L

    A. Alexakis and L. Biferale, Cascades and transitions in turbulent flows, Physics Reports767, 1 (2018)

    2018

  7. [7]

    La Porta, G

    A. La Porta, G. A. Voth, A. M. Crawford, J. Alexander, and E. Bodenschatz, Fluid particle accelerations in fully developed turbulence, Nature409, 1017 (2001)

    2001

  8. [8]

    Mordant, P

    N. Mordant, P. Metz, O. Michel, and J.-F. Pinton, Mea- surement of lagrangian velocity in fully developed turbu- lence, Physical Review Letters87, 214501 (2001)

    2001

  9. [9]

    Mordant, J

    N. Mordant, J. Delour, E. L´ eveque, A. Arn´ eodo, and J.- F. Pinton, Long time correlations in lagrangian dynam- ics: a key to intermittency in turbulence, Physical review letters89, 254502 (2002)

    2002

  10. [10]

    Yeung, Lagrangian investigations of turbulence, An- nual review of fluid mechanics34, 115 (2002)

    P. Yeung, Lagrangian investigations of turbulence, An- nual review of fluid mechanics34, 115 (2002)

    2002

  11. [11]

    Toschi and E

    F. Toschi and E. Bodenschatz, Lagrangian properties of particles in turbulence, Annual review of fluid mechanics 41, 375 (2009)

    2009

  12. [12]

    Sawford, Reynolds number effects in lagrangian stochastic models of turbulent dispersion, Physics of Flu- ids A: Fluid Dynamics3, 1577 (1991)

    B. Sawford, Reynolds number effects in lagrangian stochastic models of turbulent dispersion, Physics of Flu- ids A: Fluid Dynamics3, 1577 (1991)

    1991

  13. [13]

    S. B. Pope, Simple models of turbulent flows, Physics of Fluids23(2011)

    2011

  14. [14]

    Viggiano, J

    B. Viggiano, J. Friedrich, R. Volk, M. Bourgoin, R. B. Cal, and L. Chevillard, Modelling lagrangian velocity and acceleration in turbulent flows as infinitely differentiable stochastic processes, Journal of Fluid Mechanics900, A27 (2020)

    2020

  15. [15]

    Zamansky, Acceleration scaling and stochastic dynam- ics of a fluid particle in turbulence, Physical Review Flu- ids7, 084608 (2022)

    R. Zamansky, Acceleration scaling and stochastic dynam- ics of a fluid particle in turbulence, Physical Review Flu- ids7, 084608 (2022)

    2022

  16. [16]

    Arneodo, E

    A. Arneodo, E. Bacry, and J.-F. Muzy, Random cascades on wavelet dyadic trees, Journal of Mathematical Physics 39, 4142 (1998)

    1998

  17. [17]

    Bacry, J

    E. Bacry, J. Delour, and J.-F. Muzy, Multifractal random walk, Physical review E64, 026103 (2001)

    2001

  18. [18]

    Chevillard, C

    L. Chevillard, C. Garban, R. Rhodes, and V. Vargas, On a skewed and multifractal unidimensional random field, as a probabilistic representation of kolmogorov’s views on turbulence, inAnnales Henri Poincar´ e, Vol. 20 (Springer,

  19. [19]

    Biferale, G

    L. Biferale, G. Boffetta, A. Celani, A. Crisanti, and A. Vulpiani, Mimicking a turbulent signal: Sequen- tial multiaffine processes, Physical Review E57, R6261 (1998)

    1998

  20. [20]

    Sinhuber, J

    M. Sinhuber, J. Friedrich, R. Grauer, and M. Wilczek, Multi-level stochastic refinement for complex time se- ries and fields: a data-driven approach, New Journal of Physics23, 063063 (2021)

    2021

  21. [21]

    Friedrich, S

    J. Friedrich, S. Gallon, A. Pumir, and R. Grauer, Stochastic interpolation of sparsely sampled time series via multipoint fractional brownian bridges, Physical Re- view Letters125, 170602 (2020)

    2020

  22. [22]

    L¨ ubke, J

    J. L¨ ubke, J. Friedrich, and R. Grauer, Stochastic interpo- lation of sparsely sampled time series by a superstatistical random process and its synthesis in fourier and wavelet space, Journal of Physics: Complexity4, 015005 (2023)

    2023

  23. [23]

    T. Li, L. Biferale, F. Bonaccorso, M. A. Scarpolini, and M. Buzzicotti, Synthetic lagrangian turbulence by gen- erative diffusion models, Nature Machine Intelligence6, 393 (2024)

    2024

  24. [24]

    Guastoni and R

    L. Guastoni and R. Vinuesa, A new perspective on the simulation of stochastic problems in fluid mechanics with diffusion models: Fluid dynamics, Nature Machine Intel- ligence7, 816 (2025)

    2025

  25. [25]

    Chung, J

    H. Chung, J. Kim, M. T. Mccann, M. L. Klasky, and J. C. Ye, Diffusion posterior sampling for general noisy inverse problems, arXiv preprint arXiv:2209.14687 (2022)

    Pith/arXiv arXiv 2022

  26. [26]

    Biferale, F

    L. Biferale, F. Bonaccorso, M. Buzzicotti, and C. Calasci- betta, Turb-lagr. a database of 3d lagrangian trajectories in homogeneous and isotropic turbulence, arXiv preprint arXiv:2303.08662 (2023)

    arXiv 2023

  27. [27]

    S. G. Mallat, A theory for multiresolution signal decom- position: the wavelet representation, IEEE transactions on pattern analysis and machine intelligence11, 674 (2002)

    2002

  28. [28]

    T. Li, F. Tuteri, M. Buzzicotti, F. Bonaccorso, and L. Biferale, Deterministic diffusion models for lagrangian turbulence: Robustness and encoding of extreme events, European Journal of Mechanics-B/Fluids116, 204402 (2026)

    2026

  29. [29]

    C. K. Williams and C. E. Rasmussen,Gaussian processes for machine learning, Vol. 2 (MIT press Cambridge, MA, 2006). 6

    2006

  30. [30]

    Arn´ eodo, R

    A. Arn´ eodo, R. Benzi, J. Berg, L. Biferale, E. Bo- denschatz, A. Busse, E. Calzavarini, B. Castaing, M. Cencini, L. Chevillard,et al., Universal intermit- tent properties of particle trajectories in highly turbulent flows, Physical review letters100, 254504 (2008)

    2008

  31. [31]

    L’vov and I

    V. L’vov and I. Procaccia, Fusion rules in turbulent sys- tems with flux equilibrium, Physical review letters76, 2898 (1996)

    1996

  32. [32]

    Benzi, L

    R. Benzi, L. Biferale, and F. Toschi, Multiscale veloc- ity correlations in turbulence, Physical review letters80, 3244 (1998)

    1998

  33. [33]

    TianyiLi-92, SmartTURB/diffusion-lagr: stable (2024). Appendix A: Lagrangian dataset from Navier–Stokes simulations We use Lagrangian trajectories extracted from a di- rect numerical simulation of the three-dimensional in- compressible Navier–Stokes equations in a periodic do- main of sizeL= 2π, forced at large scales to achieve a statistically stationary...

    2024