Title: Extending Atmospheric Predictability Past 30 Days via Machine Learning

Abstract:

Conventional wisdom in atmospheric predictability suggests that deterministic weather forecasting is inherently capped at approximately two weeks, a limitation driven by the rapid escalation of errors at small spatial scales. This study contests that boundary by leveraging GraphCast, a machine-learning-based weather model, to optimize initial conditions for twice-daily forecasts covering the year 2020. The methodology resulted in an 86% average reduction in error at the ten-day mark compared to control forecasts derived from reanalysis initial conditions, with predictive skill persisting for more than 30 days. An examination of the mean optimal initial-condition perturbations uncovered large-scale, spatially coherent adjustments, largely characterized by an intensification of the Hadley circulation. Furthermore, when GraphCast-optimized initial conditions were applied within the Pangu-Weather model, a 21% error reduction was observed, with peak effectiveness occurring at four days. This cross-model success implies that the analyzed corrections address both errors inherent in the analysis and those within the model itself. These findings confirm that specific initial conditions exist capable of sustaining skillful deterministic forecasts well beyond the traditional two-week horizon. However, the feasibility of identifying such conditions in real-time to enhance operational weather forecasting is left for subsequent investigation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC