Generative emulation of chaotic dynamics with coherent prior
Juan Nathaniel, Pierre Gentine · 2025
Read Generative emulation of chaotic dynamics with coherent prior by Juan Nathaniel, Pierre Gentine (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-2504-14264.txt.
Introduction
Juan Nathaniel and Pierre Gentine present a way to mimic chaotic systems with a made data model that keeps its shape over long spans. Their paper from 2025 is called Cohesion. Mimics of such systems tend to drift into forms that could not occur in the real system.
The method first draws the large scale frame of the flow with a fast reduced model. That frame then guides a cleaning process that fills in the small scale detail. By cleaning a whole span at once, the method frames forecast as a kind of path planning. Tests run from a classic flow case to water equations and to climate spans of weeks to seasons. Long range skill beat past methods in those tests.
Catalog record
| CREATOR | Juan Nathaniel, Pierre Gentine |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-14264
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Generative emulation of chaotic dynamics with coherent prior Juan Nathaniel, Pierre Gentine Abstract Data-driven emulation of nonlinear dynamics is challenging due to long-range skill decay that often produces physically unrealistic outputs. Recent advances in generative modeling aim to address these issues by providing uncertainty quantification and correction. However, the quality of generated simulation remains heavily dependent on the choice of conditioning priors. In this work, we present an efficient generative framework for dynamics emulation, unifying principles of turbulence with diffusion-based modeling: Cohesion. Specifically, our method estimates large-scale coherent structure of the underlying dynamics as guidance during the denoising process, where small-scale fluctuation in the flow is then resolved. These coherent priors are efficiently approximated using reduced-order models, such as deep Koopman operators, that allow for rapid generation of long prior sequences while maintaining stability over extended forecasting horizon. With this gain, we can reframe forecasting as trajectory planning, a common task in reinforcement learning, where conditional denoising is performed once over entire sequences, minimizing the computational cost of autoregressive-based generative methods. Empirical evaluations on chaotic systems of increasing complexity, including Kolmogorov flow, shallow water equations, and subseasonal-to-seasonal climate dynamics, demonstrate Cohesion superior long-range forecasting skill that can efficiently generate physically-consistent simulations, even in the presence of partially-observed guidance. keywords generative modeling, diffusion, turbulence, subseasonal-to-seasonal climate 1 Introduction Solving partial differential equations (PDEs) with probabilistic emulators Gao et al. (2024) ; Rühling Cachay et al. (2024) ; Gilpin (2024) has gained significant traction over deterministic methods Brandstetter et al. (2022a) ; Brandstetter et al. (2022b) ; Li et al. (2020) ; Lu et al. (2021) because of their ability to generate ensemble forecasts, which are useful for uncertainty quantification and decision-making processes. In particular, diffusion models – a powerful class of probabilistic models – have been widely used as emulators in an autoregressive manner to produce sequential outputs over a target lead time, t\in[0,T] Li et al. (2024) ; Price et al. (2023) ; Lippe et al. (2024) . However, this approach presents several challenges. First, the generative process to estimate the posterior p(\mathbf{u}\mid\mathbf{c}) (where \mathbf{u}\in\mathbb{R}^{T\times n_{\mathbf{u}}} is the state vector and \mathbf{c}\in\mathbb{R}^{T\times n_{\mathbf{c}}} is the conditioning vector) is highly dependent on the quality of \mathbf{c} Stock et al. (2024) ; Zhao et al. (2024) ; Gong et al. (2024) ; Gao et al. (2024) ; Chen et al. (2023) ; Gao et al. (2024) ; Hua et al. (2024) ; Li et al. (2024) . Second, diffusion-based autoregressive methods are computationally expensive because they require multiple denoising steps, and the number of function evaluations (NFEs) grows with the discretization granularity of T Price et al. (2023) ; Lippe et al. (2024) . This cost becomes a bottleneck for long-range forecasting applications – such as in weather and climate – where previous gains in inference speed achieved by deterministic deep emulators are quickly offset. Figure 1 : Overview of Cohesion: an emulation framework integrating turbulence principles with diffusion-based modeling. A reduced-order model (ROM) emulates large-scale coherent flow, which serves as a conditioning prior in the generative, resolving process. Inspired by trajectory planning, Cohesion incorporates several key features to enhance flexibility and consistency, including: (a) classifier-free guidance for handling a broad range of conditioning scenarios – e.g., using different coherent estimators in a “plug-and-play” manner; (b) temporal composition to accommodate variable-horizon sequences; and (c) model-free convolution to ensure global consistency. In response to these challenges, we introduce Cohesion (Figure 1 ), a diffusion-based emulator that leverages turbulence principles to produce stable and skillful long-range rollouts, while achieving orders-of-magnitude speedups over state-of-the-art baselines. We begin by establishing a theoretical framework in which the generative process in diffusion is interpreted as a closure problem – that is, one in which the dynamics are expressed as a function of the mean flow – guided by the coherent prior \mathbf{c} . Building upon this framework, we estimate coherent prior using reduced-order models (ROMs), such as the deep Koopman operator Lusch et al. (2018) ; Wang et al. (2022) . ROMs are adept at capturing dynamics that evolve on low-dimensional attractors dominated by persistent coherent structures Stachenfeld et al. (2021) ; Solera-Rico et al. (2024) , and they offer greater stability over long rollouts compared to high-dimensional models Nathaniel et al. (2024) . Nonetheless, our framework is agnostic to the choice of coherent estimator, relying on classifier-free guidance that remains invariant to the prior as long as they are stable and effectively capture large-scale dynamics. By redefining the conditioning vector as a flow, it becomes natural to first generate the coherent trajectory and then perform a conditional pass to resolve the full dynamics in one go. This formulation parallels trajectory planning, a well-studied problem in reinforcement learning (RL) Janner et al. (2022) , and motivates several techniques to enhance the flexibility and consistency of Cohesion: (i) classifier-free guidance to support a broad range of conditioning scenarios (Figure 1 a), (ii) temporal composition to provide flexibility for variable-horizon emulation (Figure 1 b), and (iii) model-free convolution to improve spatiotemporal consistency (Figure 1 c). We evaluate Cohesion by studying chaotic spatiotemporal dynamics of increasing complexity, including Kolmogorov flow (KF), shallow water equation (SWE), and subseasonal-to-seasonal dynamics (S2S; beyond weather timescale, 6-week lead times). For instance, we demonstrate that our framework is both stable and skillful over long rollouts when compared to state-of-the-art probabilistic baselines, including ensemble forecasts from the four leading national weather agencies in the S2S case. Moreover, Cohesion exhibits minimal spectral divergences, highlighting its ability to resolve multiscale physics even in the presence of a partially observed prior. Results are presented in Section 2 , with discussion in Section 3 , and methodologies detailed in Section 4 . 2 Results 2.1 Unified Turbulence-Diffusion Framework We recast recent works on diffusion-based dynamics forecasting through the lens of turbulence theory. Specifically, we consider discrete dynamics in one temporal dimension t\in[0,T] : \mathbf{u}(t+1)=\mathcal{F}[\mathbf{u}(t)] (1) where \mathbf{u}(t)\in\mathbb{R}^{n_{\mathbf{u}}} is the state vector at time t , and \mathcal{F}:\mathbb{R}^{n_{\mathbf{u}}}\to\mathbb{R}^{n_{\mathbf{u}}} differentiable flow map. A common approach to represent these dynamics is to decompose the state \mathbf{u} into a coherent (mean) flow and a fluctuating (turbulent) component. This can be broadly captured by an operator \mathcal{H}:\mathbb{R}^{T\times n_{\mathbf{u}}}\times\mathbb{R}^{T\times n_{\mathbf{u}}}\to\mathbb{R}^{T\times n_{\mathbf{u}}} , where the simplest linear case corresponds to the classical Reynolds decomposition: \mathbf{u}=\mathcal{H}[\underbrace{\bar{\mathbf{u}}}_{\text{coherent flow}},\underbrace{\mathbf{u}^{\prime}}_{\text{fluctuating flow}}]=\underbrace{\bar{\mathbf{u}}+\mathbf{u}^{\prime}}_{\text{Reynolds decomposition}} (2) Often, the turbulent flow is expressed as a function of the mean flow, which is known as the closure problem Vinuesa and Brunton (2022) . One of the earliest approaches was proposed by Boussinesq Boussinesq (1903) , who related Reynolds stresses, \overline{u^{\prime}_{i}u^{\prime}_{j}} (with the overbar denoting time averages over the i,j -th component of \mathbf{u}^{\prime} ), to the mean flow through eddy viscosity \nu_{T} . However, this approach – and many that followed – assumes a priori model for the unresolved subgrid-scale dynamics Maeyama and Watanabe (2020) . Recent studies have implicitly leveraged this concept in a more data-driven manner, particularly via generative modeling. Here, we make this connection explicit starting with the mean flow as our coherent prior, \bar{\mathbf{u}} . Many works, for instance, approximate a variation of \bar{\mathbf{u}} using a deterministic mapping, which is then used as a conditioning prior to estimate the posterior either (i) directly over the full solution p_{\theta}(\mathbf{u}_{K}\mid\bar{\mathbf{u}}) , or (ii) indirectly over