Introduction

As climate change continues to impact our world, accurately predicting weather patterns weeks in advance has become increasingly crucial. Subseasonal forecasting, which predicts weather patterns 2 to 6 weeks ahead, remains a significant scientific challenge but is vital for sectors like agriculture, disaster preparedness, and water resource management. Conventional models, while improved a lot, strong demand for their further development to support informed decision-making across various sectors. This paper introduces the FuXi Subseasonal-to-Seasonal (FuXi-S2S) model, a state-of-the-art machine learning model that outperforms traditional forecasting models in predicting global weather patterns up to 42 days in advance.

The Challenge of Subseasonal Forecasting

Subseasonal forecasting, which covers predictions from 2 to 6 weeks, bridges the critical gap between short-term weather forecasts (up to 15 days) and longer-term climate forecasts (seasonal and beyond). Historically, this intermediate timescale has been challenging due to its reliance on both atmospheric initial conditions and surface boundary conditions, neither of which provides sufficient predictability, leaving subseasonal forecasts in a so-called predictability desert. Traditional models, such as those from the European Centre for Medium-Range Weather Forecasts (ECMWF), have made strides but still face significant limitations, especially in predicting extreme weather events.

The Innovation Behind FuXi-S2S

The FuXi-S2S model represents a significant leap forward in machine learning models for subseasonal forecasting. Trained on 72 years of daily statistics from the ECMWF ERA5 reanalysis data, our model incorporates a comprehensive suite of variables, including 5 upper-air atmospheric variables at 13 pressure levels and 11 surface variables. This extensive dataset enables FuXi-S2S to provide detailed and accurate forecasts.

A key innovation of FuXi-S2S is its use of flow-dependent perturbations in its ensemble forecasts. Traditional models often struggle with constructing initial condition perturbations due to the complexities of multivariate interactions. Our model introduces perturbations directly into the hidden features of the machine learning model, significantly enhancing forecast performance by better capturing forecast uncertainty.

Superior Performance

The performance of FuXi-S2S was evaluated using data from 2017 to 2021, comparing it with the ECMWF's state-of-the-art Subseasonal-to-Seasonal (S2S) ensemble. FuXi-S2S consistently outperforms in both deterministic and probabilistic forecasts, particularly in predictions for total precipitation and outgoing longwave radiation (OLR).

One of the most significant advantages of FuXi-S2S is its ability to predict extreme weather events. These events, such as heavy rainfall or droughts, profoundly impact societies and economies. FuXi-S2S identifies extreme precipitation events when total precipitation exceeds the 90th climatological percentile. It consistently shows higher skills for these events, particularly in extra-tropical regions and over land, where disaster preparedness is critical. A notable example of FuXi-S2S's capability is its accurate prediction of the 2020 Meiyu season in the Yangtze-Huaihe River Valley. This event saw record-breaking rainfall, leading to severe flooding. FuXi-S2S predicted this extreme event four weeks in advance, providing valuable lead time for mitigation efforts.

Furthermore, the better performance of FuXi-S2S may be attributed to its superior capabilities in predicting the Madden-Julian Oscillation (MJO), a key driver of global climate patterns. FuXi-S2S successfully extends the skillful MJO prediction from 30 days to 36 days.

Beyond Accuracy: Discovering Precursor Signals

FuXi-S2S goes beyond mere accuracy by identifying potential precursor signals, offering researchers insights into Earth systems. By generating saliency maps, the model highlights key geographic regions that significantly impact its predictions. This feature is particularly useful in understanding and validating the decision-making mechanisms of machine learning models, leading to increased trust and implementation of effective actions.

For instance, during the 2022 Pakistan floods, FuXi-S2S's predictions aligned closely with previous studies, highlighting the cooling of equatorial central Pacific and the emerging of negative Indian Ocean Dipole (IOD) event that significantly contributed to the flooding. This ability to provide actionable insights based on machine learning predictions marks a significant advancement in subseasonal forecasting.

The Future of Subseasonal Forecasting

FuXi-S2S represents a major leap forward in subseasonal forecasting. Its superior performance in prediction skills as well as its innovative ensemble prediction system, sets a new benchmark in the field. By bridging the gap between traditional models and advanced machine learning techniques, FuXi-S2S not only enhances forecast accuracy but also provides deeper insights into underlying physical processes.

As we face climate change and increasing number of extreme weather events, models like FuXi-S2S are indispensable. They offer the potential to improve disaster preparedness, optimize agricultural planning, and manage water resources more effectively. The continued development and refinement of such models will be crucial in building a more resilient future.

In conclusion, FuXi-S2S demonstrates the power of machine learning in addressing some of the most challenging aspects of weather forecasting. Its innovative approach and superior performance highlight the immense potential of integrating advanced technologies into climate science, paving the way for more accurate and reliable subseasonal forecasts.