Reduced-Order Modeling of Cyclo-Stationary Time Series Using Score-Based Generative Methods
Ludovico Theo Giorgini, Tobias Bischoff, Andre Noguiera Souza · 2025
Read Reduced-Order Modeling of Cyclo-Stationary Time Series Using Score-Based Generative Methods by Ludovico Theo Giorgini, Tobias Bischoff, Andre Noguiera Souza (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-2508-19448.txt.
Introduction
Ludovico Theo Giorgini, Tobias Bischoff, and Andre Noguiera Souza build small stand-in models for climate data that rises and falls with the seasons, in a 2025 paper. Many natural systems swing on yearly and daily cycles, and full climate models are slow to run. A lighter model that keeps the key patterns can stand in for quick tests.
Their method learns from a score-based generative model, a tool that learns the shape of the data itself. They test it on the Planet Simulator climate model, tracking the top twenty patterns of surface temperature through the year. The stand-in copies the spread, timing links, and map patterns of the real output well. Such copies let scientists run many more trials at low cost.
Catalog record
| CREATOR | Ludovico Theo Giorgini, Tobias Bischoff, Andre Noguiera Souza |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-19448
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Reduced-Order Modeling of Cyclo-Stationary Time Series Using Score-Based Generative Methods Ludovico Theo Giorgini, Tobias Bischoff, Andre Noguiera Souza Abstract Many natural systems exhibit cyclo-stationary behavior characterized by periodic forcing such as annual and diurnal cycles. We present a data-driven method leveraging recent advances in score-based generative modeling to construct reduced-order models for such cyclo-stationary time series. Our approach accurately reproduces the statistical properties and temporal correlations of the original data, enabling efficient generation of synthetic trajectories. We demonstrate the performance of the method through application to the Planet Simulator (PlaSim) climate model, constructing a reduced-order model for the 20 leading principal components of surface temperature driven by the annual cycle. The resulting surrogate model accurately reproduces the marginal and joint probability distributions, autocorrelation functions, and spatial coherence of the original climate system across multiple validation metrics. The approach offers substantial computational advantages, enabling generation of centuries of synthetic climate data in minutes compared to weeks required for equivalent full model simulations. This work opens new possibilities for efficient modeling of periodically forced systems across diverse scientific domains, providing a principled framework for balancing computational efficiency with physical fidelity in reduced-order modeling applications. 1 Introduction Many natural systems, from fluid dynamics to climate science, are characterized by interactions across a vast hierarchy of temporal and spatial scales, making their direct numerical simulation computationally prohibitive [ 10 ] . This challenge has motivated decades of research into reduced-order stochastic models that aim to capture the essential macroscopic behavior of the full dynamics within a tractable mathematical framework [ 35 , 49 , 24 , 24 , 46 , 40 , 33 , 57 , 12 , 13 ] . The foundational work of Hasselmann [ 29 ] established stochastic climate theory by framing slow climate variability as the integrated response to fast ”weather” noise, inaugurating a rich tradition of energy-balance models with feedbacks and memory kernels that yield realistic spectral characteristics and finite predictability horizons. A central challenge in model reduction is ensuring that the simplified model remains physically consistent, particularly by preserving the long-term statistical properties, or invariant measure, of the original system. Majda’s systematic approach to stochastic mode-reduction established rigorous theoretical foundations for slow-fast coarse-graining in nonlinear geophysical systems [ 41 ] , while empirical linear inverse models (LIMs) have provided practical routes to reduced operators capturing coupled atmosphere-ocean dynamics [ 47 , 2 , 45 ] . Modern advances have extended these frameworks to handle non-Gaussian statistics and extreme-value properties, with recent work deploying barotropic and triad prototypes to show that stochastic reductions can preserve both extreme-event statistics and predictability envelopes [ 9 ] . In parallel, data-driven response-operator methods and practical GFDT implementations have emerged in nonlinear settings [ 25 , 22 , 8 , 5 , 18 , 43 , 42 , 27 , 17 , 7 , 28 , 6 , 11 ] . A recent breakthrough in this field was the development of a data-driven framework that addresses the physical consistency challenge for autonomous (time-invariant) systems through explicit decomposition of the system’s drift into conservative and non-conservative components [ 19 ] . This approach leverages advances in score function estimation—the gradient of the logarithm of the steady-state probability density—combined with temporal correlation statistics to solve for a drift matrix that preserves both the invariant measure and short-term dynamics. The score matching methodology, originally developed by Hyvärinen [ 31 ] for exponential-family models, has been revolutionized by denoising approaches [ 32 , 56 ] that connect score estimation to conditional noise prediction, enabling robust density gradient estimation in high-dimensional spaces without requiring explicit normalization constants. However, many critical systems are not autonomous. In climate science, the Earth’s dynamics are fundamentally cyclo-stationary, driven by the strong periodic forcing of the annual and diurnal cycles [ 44 ] . Such systems present unique challenges that have motivated the development of specialized statistical frameworks. Cyclo-stationary linear inverse models (CS-LIM) explicitly incorporate seasonally varying operators and stochastic forcing, demonstrating superior performance over stationary counterparts in ENSO and marine heatwave prediction [ 1 , 24 ] . Recent work has also explored hierarchical spatio-temporal Gaussian-process frameworks that use circular time embeddings to emulate annual cycles over large climate datasets [ 3 ] , while fractional and non-Markovian energy-balance models introduce power-law kernels that reproduce slow seasonal lag structures [ 36 ] . Related emulator and representation frameworks further advance statistical modeling of climate fields and spatial structure [ 53 , 16 , 20 , 52 , 23 ] . The mathematical foundation for decomposing dynamics into conservative and non-conservative components has deep roots in the theory of stochastic differential equations. Ao’s seminal work [ 4 ] on irreversible diffusion splitting demonstrated that any Fokker-Planck drift can be partitioned into gradient (potential) and divergence-free circulatory components, providing a rigorous framework for separating reversible and irreversible dynamics. In linear Ornstein-Uhlenbeck systems, antisymmetric drift components encapsulate irreversible ”circulation” that generalizes Hamiltonian dynamics, while recent neural approaches have enabled simultaneous learning of Hamiltonian and Rayleigh dissipation terms in complex flows [ 34 ] . From the response-theoretic viewpoint, generator-based fluctuation–dissipation formulations offer interpretable routes to prediction and control in nonequilibrium systems [ 38 , 37 ] . Despite these advances, several research gaps remain. Neural denoising score matching has achieved remarkable success primarily in Euclidean latent spaces, but performance on manifolds or augmented phase spaces requires Riemannian generalizations [ 50 ] . Most cyclo-stationary treatments remain linear, and extending conservative-irreversible decomposition to nonlinear, periodically forced regimes represents an open frontier [ 51 ] . In this paper, we extend the data-driven decomposition framework to cyclo-stationary systems by addressing these limitations through a novel state-space augmentation approach. The core idea is to recast the non-autonomous problem into an autonomous one in a higher-dimensional state space. This is achieved by augmenting the state space with periodic functions (e.g., sine and cosine of the annual cycle) to make the system’s state explicitly dependent on time, and constructing a single, constant drift matrix for this augmented system, which implicitly encodes the time-dependent dynamics of the original physical variables while preserving the conservative-non-conservative decomposition. This approach bypasses the need for time-dependent parameters by increasing the dimensionality of the state space, embedding the temporal dependence into the structure of the joint probability distribution. We demonstrate the power of this extended framework by applying it to data from the PlaSim (Planet Simulator) climate model, building a reduced-order model for the 20 leading principal components of the surface temperature field driven by the annual cycle. Our results show that the resulting surrogate model accurately reproduces both the marginal probability distributions and autocorrelation functions of the principal components. Notably, this work represents the first application of the approach proposed in [ 19 ] to a relatively high-dimensional system; in the cited paper, the method was tested on systems with up to only three dimensions. This paper, therefore, demonstrates the relevance and scalability of the framework, validating its efficacy for modeling complex, periodically forced systems and opening promising avenues for climate projection, extreme event analysis, and understanding critical transitions in natural systems subject to periodic forcing. Section 2 revisits the data-driven decomposition framework for autonomous dynamical systems, laying the mathematical groundwork for the present extension.