Physics-aware generative models for turbulent fluid flows through energy-consistent stochastic interpolants

Nikolaj T. Mücke, Benjamin Sanderse · 2025

Read Physics-aware generative models for turbulent fluid flows through energy-consistent stochastic interpolants by Nikolaj T. Mücke, Benjamin Sanderse (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-05852.txt.

Introduction

Nikolaj Mücke and Benjamin Sanderse build a generative model for turbulent fluid flow in this 2025 paper. Turbulence tests are costly to run with classic solvers. Generative models can learn to produce flow fields, but most ignore the physics of the flow. Their model rests on stochastic interpolants, which blend noise into data along a random path. They make parts of that path learnable, so the model can enforce energy stability and flow without sources or sinks.

They test the model on Kolmogorov flow, a standard turbulence benchmark. It beat strong rivals, among them a diffusion based forecaster and a refinement method, on accuracy and stability. It also stayed stable over far longer roll outs than standard interpolants. The authors present physics aware generation as a way to speed turbulence work without dropping core conservation laws. The method points toward cheaper simulation for engineering and weather related tasks.

Catalog record

CREATORNikolaj T. Mücke, Benjamin Sanderse
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE78 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-05852

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-aware generative models for turbulent fluid flows through energy-consistent stochastic interpolants Nikolaj T. Mücke, Benjamin Sanderse Abstract Generative models have demonstrated remarkable success in domains such as text, image, and video synthesis. In this work, we explore the application of generative models to fluid dynamics, specifically for turbulence simulation, where classical numerical solvers are computationally expensive. We propose a novel stochastic generative model based on stochastic interpolants, which enables probabilistic forecasting while incorporating physical constraints such as energy stability and divergence-freeness. Unlike conventional stochastic generative models, which are often agnostic to underlying physical laws, our approach embeds energy consistency by making the parameters of the stochastic interpolant learnable coefficients. We evaluate our method on a benchmark turbulence problem – Kolmogorov flow – demonstrating superior accuracy and stability over state-of-the-art alternatives such as autoregressive conditional diffusion models (ACDMs) and PDE-Refiner. Furthermore, we achieve stable results for significantly longer roll-outs than standard stochastic interpolants. Our results highlight the potential of physics-aware generative models in accelerating and enhancing turbulence simulations while preserving fundamental conservation properties. Journal: Computers and Fluids 1 Introduction Recently developed AI tools like ChatGPT [ 1 ] , Sora [ 2 ] , and DALL-E [ 3 ] mark a ground-breaking era in generative modelling across text, video, and image domains. A natural question is whether generative models can also be used to support or replace simulation codes that have been classically used to model complex problems arising in domains like physics, chemistry, or biology. In particular, our interest lies in generative models for fluid dynamics problems. Fluid dynamics simulation codes have been developed in the last decades based on known physical principles (e.g. conservation laws), but with the main limitation that they are computationally typically very expensive to run, and also very difficult to maintain. Pre-trained generative models hold the promise that they can significantly accelerate physics simulations while at the same time providing a notion of uncertainty in the outcome, and having the ability to model unresolved scale processes [ 4 ] . A first step in this direction are the “foundation models” that have been developed for weather prediction [ 5 , 6 , 7 , 8 ] . These foundation models [ 9 ] are very large neural networks (sometimes having more than 1 billion parameters [ 5 ] ), that are pre-trained on vast (re-analysis) datasets, and can be finetuned in space and time to specific tasks. These models have shown the potential to accelerate classical numerical weather predictions to the point that a 5-day forecast can be made in less than a minute [ 5 ] . The foundation model approach has also been applied to computational chemistry [ 10 ] , biology [ 11 ] , and fluid mechanics [ 12 ] . In [ 12 ] , a foundation model was pre-trained on certain PDEs (compressible Euler and incompressible Navier-Stokes) and gave accurate results on PDEs that were not in the pre-training set. These foundation models are pre-trained, but unlike their counterparts in text, video and image generation, they are still usually deterministic in nature and mainly focus on forecasting the mean of the possible trajectories [ 5 , 8 , 12 , 13 ] . This can lead to blurred forecast states [ 8 , 14 ] . A few stochastic machine learning approaches have recently been introduced to address this issue [ 15 , 13 , 16 ] . More in general, stochasticity naturally arises when developing reduced models of fluid dynamics problems, in which the state is typically split into large (resolved) and small (unresolved) scales, and the purpose is to infer models for only the large scales. An example is through the Mori-Zwanzig formalism, in which the effect of the initial condition of the unresolved scales is typically modelled by a noise term. Similarly, one can argue that the evolution of the large scales is not deterministic, even when the evolution of the full model is deterministic. This manifests itself in non-uniqueness of the problem (different small scale realizations can have the same effect on the large scales), which warrants the use of a stochastic approach [ 17 ] . We note here that, outside the realm of machine learning, stochastic approaches have been long in use to model unresolved processes, typically known as ‘stochastic parameterizations’ [ 4 , 18 ] . The stochastic approach that we are considering is in line with those approaches, but benefits from the advanced approximation capabilities of neural networks. Previous research has explored the approximation of SDEs using neural networks. In [ 19 ] neural networks are trained with gradients computed via adjoint methods. While such methods alleviate some of the computational burden compared with naively backpropagating through a solver, it is still prohibitively expensive for very high-dimensional problems. Specifically, systems of up to 50 dimensions are studied in [ 19 ] . In [ 20 ] similar approaches are utilized for learning a neural SDE as a closure term for LES simulations. To minimize the computational restrictions of high-dimensional systems, the discrete state dimension is reduced via an autoencoder and the SDE is trained in the latent space. While this approach is indeed promising, the use of autoencoders makes it difficult to infuse physics knowledge into the model due to the non-physical nature of the latent space. Furthermore, this approach relies on an LES solver, as the neural network only approximates a closure term. It is unclear how this approach would fare as a full non-intrusive surrogate model. In this work, as an alternative to the expensive adjoint-based training methods, we will draw inspiration from the field of generative models. The current state-of-the-art in stochastic generative models for fluid flows relies mainly on denoising diffusion probabilistic models (DDPMs) [ 21 , 13 ] – similar to what is used in machine learning models for image generation [ 22 , 23 ] . DDPMs are based on sampling from a Gaussian distribution and transforming it through a sequence of steps into a realistic image. In the context of time-dependent problems, one deals with a sequence of time steps, and a realistic ‘image’ (e.g. a flow field) needs to be created at each time step . This significantly complicates the process compared to image generation. An example of how this can be achieved is given in [ 13 ] : sample from a Gaussian at each time step, denoise, and condition on the state at the previous time step. A similar approach is used in [ 21 ] . Denoising at each time step is rather involved and not in line with the actual physical process. In general, a common issue with existing stochastic generative models is that they are agnostic of the underlying physical processes that are being modelled. For example, conservation of mass and energy are fundamental physical principles in fluid flows, but are not obeyed by existing generative models. This can lead to issues with the stability of predictions, especially when considering long prediction horizons. In this work, we develop a pre-trained, stochastic generative model that (to the authors’ knowledge) incorporates for the first time a sense of physical awareness. Our approach is built upon two key components: stochastic interpolants

Read the complete text