SDECast uses neural SDEs for continuous-time probabilistic weather forecasting, recovering drift dynamics in a simulated geophysical flow and producing skillful global hourly forecasts up to five days
Related research and updates
Synopsis
The authors introduce SDECast, a Neural Stochastic Differential Equation (SDE) framework for continuous-time probabilistic weather forecasting that extends SDE Matching to learn stochastic dynamics directly in physical space without repeated SDE simulation during training; on a simulated geophysical flow, SDECast recovers meaningful drift dynamics and faithfully reproduces the underlying continuous-time behavior, and it then scales to global weather forecasting at hourly resolution, producing skillful probabilistic forecasts for lead times up to five days.
Figure 1 : The RMSE ( ↓ \downarrow ), CRPS ( ↓ \downarrow ), SSR for experiments on SQG.
arXivInterpretation
SDECast formulates weather forecasting as a neural stochastic differential equation in continuous time, learning stochastic dynamics directly in physical space rather than through autoregressive rollouts at a fixed temporal resolution. Existing machine learning weather forecasting models typically generate forecasts through autoregressive rollouts at a fixed temporal resolution, which can suffer from severe error accumulation with shorter time steps and does not explicitly encode the locality and temporal continuity of atmospheric dynamics; SDECast responds to both points with a continuous-time SDE formulation. The abstract states that the framework extends SDE Matching and does not require repeated SDE simulation during training, which is a method-level design statement.
On a simulated geophysical flow, SDECast recovers meaningful drift dynamics and faithfully reproduces the underlying continuous-time behavior. This provides a controlled simulated setting for checking whether the model learns continuous-time dynamics rather than only fitting discrete time points. The abstract reports results on a simulated geophysical flow but gives no specific error metrics, sample sizes, or control settings.
SDECast scales to global weather forecasting at hourly resolution and produces skillful probabilistic forecasts for lead times up to five days. This demonstrates scalability from a simulated flow to the global scale and advances continuous-time probabilistic forecasting to hourly resolution and a five-day lead time. The abstract reports scalability at global hourly resolution and skillful probabilistic forecasts within a five-day lead time, but gives no scoring metrics, baseline comparisons, or numerical values.
Perspective
This work targets researchers and practitioners who need continuous-time, probabilistic weather forecasting: it is used on a simulated geophysical flow to check the recovery of drift dynamics and continuous-time behavior, and at global hourly resolution for probabilistic forecasts up to five days. It applies to settings that treat atmospheric dynamics as a continuous-time stochastic process and highlights the efficiency feature of not requiring repeated SDE simulation during training.
The abstract does not give scoring metrics for the probabilistic forecasts, baseline models, forecast variables, or spatial resolution, nor does it describe the specific setup and control conditions of the simulated geophysical flow; the claim of skill at a five-day lead time lacks comparable numerical support. In addition, what is available here is the abstract rather than the full paper, so evidence in figures and supplementary materials cannot be assessed here; these are open questions for readers judging empirical strength.