Fine-tuning a global weather model for improved subseasonal forecasting

Accurate subseasonal forecasting is socio-economically critical yet remains a great scientific challenge. Recent advances in machine-learning based global weather forecasting demonstrate superior skill on medium-range (1 to 15 days ahead) and subseasonal-range (15 to 42 days ahead) than the best traditional numerical weather forecasting system. These data-driven models require immense computational resources for training, which are not widely available. Here we show, by using medium-range Graphcast model as pre-trained model and focusing on reducing iterative error accumulation as well as producing weekly mean instead of daily forecasts, that fine-tuning is an efficient strategy to achieve improved results for subseasonal forecasting [1]. Furthermore, we also implemented an 11-member ensemble forecast using different versions of GraphFT, enabling us to outperform the leading probabilistic subseasonal forecasting system (S2S ensemble) at 3-4 week leads. Our fine-tuned models rapidly converge (trained on just three years of data) and with significantly improved results at subseasonal timescales. Demonstrating the potential of fine-tuning for improving, with low computational costs, possibly both atmosphere and ocean forecasts at long leads. [1] V. Sansine, T. Izumo, M. Hopuare, D. Specq, et S. M.-L. Pierre, « Fine-tuning a global weather model for superior subseasonal forecasting », 10 mars 2025, Research Square. doi: 10.21203/rs.3.rs-5619528/v1.