Year: 2026
Type: preprint
Abstract: Subseasonal forecasting (15 to 42 days ahead) is socio-economically critical yet remains challenging. Recently, machine-learning based global weather forecasting could show superior skill than traditional forecasting systems. Yet, training these data-driven AI models require immense computational resources. Here we demonstrate, by using a medium-range model, GraphCast, as pre-trained model and focusing on reducing its iterative error accumulation, that fine-tuning is a cost-effective strategy to improve subseasonal forecasting. We further introduce a simple and computationally-inexpensive ensemble method: creating 11 variants of our fine-tuned model, FT, through different data sampling during fine-tuning and considering each as a member of a new model, FT ensemble. This model significantly surpasses the ECMWF Subseasonal to Seasonal ensemble model (S2S ensemble) for week 3 and week 4 lead-times, notably for precipitation, crucial for water management. These results highlight that combining the proposed fine-tuning and ensemble methods are powerful, low-cost pathways to improve atmospheric and ocean forecasting.
Source: Research Square
Language: English
Subfield: Atmospheric Science
Field: Earth and Planetary Sciences
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7213612289