Bridging the "Predictability Desert": A Probabilistic Bias Correction Framework for AI and Dynamical Subseasonal Forecasts

Subseasonal weather prediction (2–6 weeks lead time) represents a critical "predictability desert" where the influence of atmospheric initial conditions diminishes and boundary forcings have not yet become dominant. Despite its inherent difficulty, skillful subseasonal forecasting is vital for decision-making in agriculture, water resource management, public health and disaster preparedness. While recent breakthroughs in artificial intelligence (AI) and machine learning (ML) have revolutionized synoptic-scale weather forecasting, these gains have not yet fully translated to the subseasonal regime, where systematic biases in both dynamical and data-driven models remain a primary bottleneck.

In this work, we present a novel Probabilistic Bias Correction (PBC) framework that leverages ML to systematically identify and correct errors in global model forecasts. Our approach is model-agnostic, as we demonstrate by showing it can enhance both traditional physics-based dynamical ensembles and emerging AI-based forecasting systems. By training on historical reanalysis and model forecast datasets, the PBC framework significantly reduces systematic errors that typically degrade raw model skill at subseasonal lead times.

We evaluate the performance of our PBC algorithms against several high-standard benchmarks, including climatology, multi-model super-ensembles from major operational centers, and state-of-the-art AI models. Notably, our framework was benchmarked within the context of the AI Weather Quest (sponsored by ECMWF). Results demonstrate that our PBC forecasts outperform all participating dynamical and ML models, including the ECMWF Integrated Forecasting System (IFS) and Artificial Intelligence Integrated Forecasting System (AIFS), in predicting 2-meter temperature, precipitation and mean sea level pressure.

To demonstrate the real-world utility of this system for early warning capabilities, we present case studies of extreme winter weather events in the Eastern United States and Europe. Our model successfully predicted these high-impact events several weeks in advance, with forecasts disseminated in real-time to stakeholders via social media. Our findings suggest that while AI-based models like FuXi-S2S offer a strong alternative to dynamical systems, the integration of probabilistic post-processing is critical to maximize predictive skill and provide reliable, sector-specific decision support in a changing climate.