Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.03106

S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales

Chenyu Dong·Gianmarco Mengaldo
View original source ↗Read source document ↗Export metadata ↓

Abstract

The subseasonal-to-seasonal (S2S) timescale, roughly from two weeks to two months ahead, is a critical forecast window for sectors such as agriculture, energy, and water management. Yet, it is widely known as the `predictability desert'. Recent AI weather models excel up to two weeks ahead but deteriorate beyond, largely because they are trained to predict fine-scale details that are neither predictable nor essential at S2S timescales. We argue that a more physically grounded objective is to forecast only the slowly varying components that remain predictable. Computer vision reached the same conclusion with the Joint-Embedding Predictive Architecture (JEPA), which predicts in latent space, discarding unpredictable details. In this work, we introduce S2S-JEPA, which brings the JEPA paradigm to S2S forecasting. It is tailored to this task through design elements from state-of-the-art AI weather models. S2S-JEPA achieves comparable skill to the gold-standard ECMWF physics-based ensemble and surpasses it on multiple metrics at weeks 5 to 6.

Explore related subjects

physics.ao-phcs.AIcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Chenyu Dong, Gianmarco Mengaldo. 2026-10-02. S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales. https://arxiv.org/abs/2610.03106

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Baseline Climatology of Tornadoes and Waterspouts in the Philippine Archipelago

Tornadoes are documented on every inhabited continent, yet the environmental conditions used to anticipate them were developed and validated overwhelmingly on mid-latitude records, leaving tropical regimes largely unexamined. Here, we collected 1,186 combined tornado and waterspout reports in the Philippines spanning the early 1900s to 2025 and paired each with an ERA5 proximity sounding to construct the first baseline climatology and environmental characterization of these events in the archipelago. Activity concentrates in three corridors, mostly flanked by mountain ranges, one in each major island group, and is sharply phased within a May-October window that closely tracks the annual cycle of thunderstorm frequency, peaking in August and lagging into October for waterspouts alongside the tropical cyclone season. These events cluster in the mid-to-late afternoon, consistent with a daytime heating regime rather than nocturnal convection. The environments are thermodynamically permissive but kinematically shallow, with convective available potential energy (CAPE) exceeding typical Great Plains values, while the 0-6 km bulk wind difference (BWD; DLS) remains weak at $\sim$10-14 kt, departing from the mid-latitude covariance in which instability and shear increase together. Although streamwise vorticity is modest in magnitude, nearly all discriminating variance resides in the lowest kilometer, where 68-79% of near-surface horizontal vorticity is streamwise, reaching $>$0.005 s$^{-1}$ and 96% streamwiseness in the most organized profiles, and the hodographs retain low-level curvature. Consequently, composite indices remain low even as tornadoes and waterspouts recur. Mid-latitude-calibrated parameters can therefore misdiagnose tropical-maritime regimes, and recalibration toward shallow-layer, orientation-aware kinematic discriminators may recover the diagnostic signal.

physics.ao-ph↗
arXiv2026

Stochastic thermodynamics and predictability of Ornstein-Uhlenbeck processes: an application to the Madden-Julian Oscillation

Climate oscillations are intrinsically out-of-equilibrium phenomena, yet the connection between their predictability and nonequilibrium thermodynamic properties remains unexplored. We develop a coarse-grained stochastic thermodynamic framework for Ornstein-Uhlenbeck (OU) processes and apply this framework to climate data via linear inverse models (LIMs), linking predictability, entropy production, dynamical activity, and path action. Climate oscillations are often studied in a 2d phase space of climate indices, which are even under time-reversal, motivating our focus on 2d OU dynamics with even variables. We find that coordinate-invariant properties of 2d OU processes can be reduced to a minimal three-parameter description: two diffusivities and a phase-space rotation. We show that these quantities fundamentally characterize distinct aspects of the dynamics. We find that higher irreversibility and dynamical activity do not necessarily imply reduced predictability. Rather, it is the path action which is most closely connected to a commonly used measure of predictability, the anomaly correlation coefficient (ACC). As an illustrative example, we apply this framework to the Madden-Julian Oscillation (MJO) over the twentieth century. We find that the MJO has become simultaneously more predictable, as seen previously, while at the same time becoming more irreversible and more active. Predictability is driven predominantly by the reduced mean diffusion. Entropy production reflects the combined evolution of all three parameters. These results suggest that stochastic thermodynamics provides a powerful lens for studying nonequilibrium climate phenomena.

physics.ao-ph↗
arXiv2026

A Multivariable Approach to the Regional Calibration of the Galvez Davison Index across Mexico Climate Regimes

The Gálvez Davison Index (GDI) is a thermodynamic diagnostic designed to summarize convective favorability in the tropical regions, yet its transferability across complex terrain and heterogeneous hydroclimates remains uncertain. Here we assess the climatological behavior and precipitation relevance of the GDI over Mexico using ERA5 hourly fields and CHIRPS daily precipitation for 1981 to 2024. We characterize the seasonal cycle and spatial contrasts of GDI and evaluate its association with rainfall using grid point and regional Pearson and Spearman correlations between daily maximum GDI and daily precipitation. To introduce an independent convective proxy, we also quantify the relationship between hourly GDI and ERA5 outgoing longwave radiation (OLR). Event level discrimination skill is assessed with receiver operating characteristic (ROC) curves and area under the curve (AUC) across regions and months. While GDI precipitation correlations are moderate and region dependent, GDI OLR coupling is consistently stronger, partly because OLR approaches a lower bound once deep convection develops, supporting its use as an intermediate proxy between thermodynamic favorability and rainfall. However, GDI values are systematically lower over elevated terrain because of the pressure level formulation of the index, which should be considered when interpreting the GDI over mountainous regions. ROC results indicate skill above random chance in nearly all region month combinations, with AUC frequently exceeding 0.75 during the core convective season. Building on these diagnostics, we propose a regionally calibrated GDI convective regime scale for May to October, adapted from the original GDI framework but recalibrated for Mexico using event conditioned distributions defined from OLR and precipitation and quantified with bootstrap uncertainty.

physics.ao-ph↗