Gappy Reconstruction of Bubbly Flows by Guided Diffusion Models

arXiv:2606.29843 · physics.flu-dyn, cond-mat.soft · Submitted 2026-06-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Gappy Reconstruction of Bubbly Flows by Guided Diffusion Models".

Paper summary: Kai: So, wrapping up, the paper "Gappy Reconstruction of Bubbly Flows by Guided Diffusion Models" shows that using a guided diffusion model can reconstruct the velocity field inside bubbles even when you only have phase limited measurements.

Mira: It’s about taking that challenge—the substantially different velocity statistics across phases—and using this specific machine learning framework to generate physically realistic fields that match what direct numerical simulations predict.

Lev: For me, the main implication is that we can use these statistical inferences to get a better starting point for other complex simulations or error correction algorithms where we don't have perfect data coverage.

Kai: And they also did show temporal reconstruction, which suggests this isn't just about spatial snapshots; it can handle gappy time signals too, even with different statistics in the masked regions.

Mira: So the authors demonstrate that this method is feasible for reconstructing bubbly flows by using diffusion models, and it’s relevant because their results are statistical and directly applicable to experimental setups.

Lev: The limitation they pointed out is that continuity across slices isn't guaranteed inherently, so any real application would need to build in a way to enforce those physical constraints.

Kai: That's the practical reality—it’s a powerful tool for inference, but we still have work to do on making it physically exact for engineering purposes.

Conclusion: Kai: So, we're talking about "Gappy Reconstruction of Bubbly Flows by Guided Diffusion Models," and what this paper is really doing is showing you how to figure out what's happening inside a bubble even when your measurements are patchy.

Mira: Right, it takes those scattered measurements and uses these diffusion models to generate a full picture of the flow inside the bubble, which is pretty impressive given how different the speeds are between the liquid and the gas parts.

Lev: From an engineering standpoint, it’s cool because you don't need perfect data coverage for every single point to get a reasonable reconstruction, as long as you have some guidance in that measured area.

Kai: Exactly, they show that this guided diffusion model can actually reproduce the 3d flow structures inside the bubble pretty well once you assemble the 2d fields they reconstruct <ref:2606.29843#pg3>.

Mira: But we need to remember, it's a statistical reconstruction, so we're talking about a probability distribution of what *could* be there, not necessarily a perfectly deterministic physical field yet.

Lev: That’s the caveat I always keep in mind; continuity across those boundaries isn't guaranteed by the model itself, so you’d need to add extra constraints for real hardware.

Kai: Yeah, they did mention that the method is statistical in nature, and that any future work should involve explicitly adding those bubble-liquid interface conditions and the incompressibility rules into the diffusion model structure.

Mira: That makes sense because right now, it’s a powerful tool for inference on these complex systems, but it’s not yet a complete physics engine for bubbly flows.

Lev: So what this means for us in the long run is that we have a way to get a better initial guess or fill in data gaps when we are trying to simulate these kinds of multiphase dynamics.

Kai: And that brings us right back to how they handled the temporal reconstruction, because they showed it works even when you have different statistics in the masked region compared to the unmasked part.

Mira: That’s a neat extension, showing this framework isn't just for static pictures but can handle time-series data too.

Lev: So next up, we look at how this reconstruction method could actually be used to improve error correction techniques for real hardware simulations of these flows.

Hridey Narula, *Tianyi Li, *Michele Buzzicotti, *Luca Biferale, Prasad Perlekar

Tata Institute of Fundamental Research · Department of Physics and INFN, University of Rome “Tor Vergata”

physics.flu-dyn, cond-mat.soft

Submitted: 2026-06-29

Updated: 2026-11-16

Comments: accepted in PRX-Intelligence, updated, includes supplementary material, 16 pages main text, 5 pages supplementary material

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 75/100

The gist: Flow reconstruction from phase-limited measurements is a challenging problem due to the substantially different velocity statistics across the phases.

Key concepts

Pseudo-turbulence (PT)
This describes a state where fluid dynamics exhibit characteristics between laminar and fully turbulent flow. In bubbly flows, this occurs at moderate Galilei numbers, leading to specific energy scaling laws like k-3 or k-5/3, making the flow numerically challenging.
Diffusion Model Framework
A type of generative AI that learns complex data distributions by gradually adding noise (forward process) and then learning how to reverse that noise (backward process). It is trained to accurately predict the original clean flow configuration from a noisy version.
Guided Diffusion Model
A specialized diffusion model where the reverse denoising process is modified to include a guidance force in a specific region of interest, like the bubble phase. This allows for targeted reconstruction of flow details within that region, improving accuracy over standard methods.

Terminology

Summary

Flow reconstruction from phase-limited measurements is a challenging problem due to the substantially different velocity statistics across the phases. p. 0 The authors address this problem for buoyancy-driven bubbly flows in the pseudo-turbulence regime by using a guided diffusion model. p. 0 The reconstructed bubble-phase velocity field accurately reproduces key statistical features of the flow. p. 0

Introduction to Flow Challenges

Flows generated by rising columns of bubbles are encountered in a wide range of natural and industrial settings. The collective dynamics of a large number of bubbles appears to have some universal properties. p. 0 The complex interplay of inertia, surface tension, and viscous dissipation leads to pseudo-turbulence (PT). p. 0 At moderate Galilei numbers (Ga ≲ 1000), the energy E(k) contained in the wavenumber k has been shown to obey the scaling k −3. p. 0 More recently, studies have shown the emergence of the Kolmogorov scaling k −5/3 at sufficiently large Galilei numbers (Ga ≳ 1000). p. 0 The multiphase and multiscale nature of the bubbly flow makes their numerical and experimental investigations extremely challenging. p. 0

How the training signal is built

The training dataset is constructed from fully resolved three-dimensional (3d) direct numerical simulations (DNS) of buoyancy-driven bubbly flows in the pseudo-turbulent regime (Ga ≈ 600). The authors perform DNS on a triply periodic cubic domain with side length L = 2π discretized with N3 = 512 cubed equispaced collocation points. p. 1 The simulation domain is initialized with 12 randomly placed spherical bubbles of equal diameter D with the gas volume fraction 1/V Rϕ dx = 3. p. 12%. The Kolmogorov timescale is defined as τη = p/ν/ε. p. 1 The total number of two-dimensional samples obtained from the saved three-dimensional snapshots is 10,630. p. 1

Diffusion Model Framework

Diffusion models are a class of generative probabilistic models that learn a high-dimensional data distribution through a prescribed forward noising process and a learned reverse (backward) denoising process. The forward process transforms the clean flow configuration X0 into noisy states Xt through a Gaussian Markov chain. p. 2 The backward process approximates the inverse of the forward noising process by denoising from Xt to Xt−1 at each diffusion step. p. 2 The model is trained using the standard DDPM noiseprediction objective, minimizing L = EX0,t,ϵ h ϵ − ϵθ(Xt, t)2 i. p. 2

Reconstruction and Guidance

The authors use a training-free guidance scheme called diffusion posterior sampling (DPS) to generate samples conditioned on the measurement region M. The reverse process is modified to include a guidance force in the measured region M, as shown in Eq. p. 3(9). The optimal guidance strength ζ opt is identified as 800 which minimizes the mean square error in the bubble phase. p. 5 The reconstruction procedure substantially outperforms a smooth spline extrapolation approach. p. 8

Reconstruction Capabilities

The authors demonstrate that for a prescribed bubble shape, the threedimensional flow structures inside the bubble can be recovered by assembling the reconstructed 2d flow fields. The resulting velocity fields are nearly indistinguishable from the DNS data, supporting the conclusion that any potential memorization effects are not significant in the present analysis. p. 4 The model not only reproduces the large-scale flow patterns, but also captures the statistics of the velocity field, its gradients, and two-point correlations with high accuracy. p. 8

Temporal Reconstruction

Diffusion models can also be used for temporal reconstruction of velocity signals. The model is trained on two channels: the indicator function ϕ and the vertical velocity uz. p. 3 The temporal spectrum of the reconstructed signal agrees quite well with the ground truth spectrum. p. 4 This demonstrates that diffusion models can also be used for reconstruction of gappy time signals, even when the statistics of the masked region (bubble) is different from the unmasked region (liquid). p. 0

Conclusion

In conclusion, we demonstrate for the first time the feasibility of reconstructing the flow within the bubble phase using a guided diffusion model. The methodology and findings are of direct experimental relevance. p. 9 A natural extension of the present work would involve explicitly incorporating the bubble–liquid interfacial boundary conditions and the incompressibility constraint into the diffusion model. p. 9 The resulting 3D bubble-phase flow closely reproduces the ground truth. p. 9 The reconstruction is statistical in nature, continuity of the velocity field across slices is not inherently guaranteed. p. 9

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI system can unconditionally generate two-dimensional slices of three-dimensional buoyancy-driven bubbly flows using a guided diffusion model trained on fully resolved three-dimensional direct numerical simulations to serve as generative priors for threedimensional turbulent multiphase flows.

  2. The improved system can reconstruct the velocity field inside bubbles by conditioning the reconstruction on the surrounding liquid flow information, accurately reproducing key statistical features of the flow and outperforming naive interpolation approaches in capturing flow structures.

  3. The system can perform three-dimensional flow reconstruction by stacking neighboring two-dimensional reconstructed slices and averaging multiple candidates to mitigate lack of continuity in the out-of-plane direction, resulting in a smoother, more coherent flow field that closely reproduces ground truth statistics.

  4. The improved system can reconstruct Eulerian time signals u(x0, t) during bubble presence by training a diffusion model on indicator function and vertical velocity uz to infer the signal given measurements at all other times, successfully reconstructing gappy time signals even when phase statistics differ significantly.

Sources

Related papers