Fine-Tuning AI Foundation Models for Subseasonal Weather Forecasting

The rapid growth of weather-dependent renewable energy increases Europe’s vulnerability to grid volatility, making accurate seasonal-to-subseasonal (S2S) forecasts increasingly important. Traditional Numerical Weather Prediction systems, however, often struggle to maintain useful accuracy beyond a 10-day horizon. To address this predictability gap, we leverage recent advancements in open-source AI Foundation Models (FoMos), since they offer major advantages in computational efficiency, scalability, and adaptability compared to traditional approaches.

In this contribution we focus on demonstrating the potential of fine-tuning ECMWF’s AIFS (Artificial Intelligence Forecasting System) for improved prediction of European Weather Regimes. AIFS-ENS offers particurlarly high skill in large-scale flow. By attaching a second Decoder (called the ‘diagnostic head’) to the Processor, we demonstrate a lightweight fine-tuning strategy. The diagnostic head can be swapped out for different targets, offering flexibility to let any AI model focus on a specific region, variable(s), lead-time, and extremes. In addition, it allows for add additional input features which are relevant for the forecasting task. In this case it can allow for more explicit information on the MJO dynamics which are important for the modeling the Weather Regimes. By focusing on large-scale regime-level dynamics, we show how AI models can better exploit sources of long-range predictability compared to traditional approaches. We evaluate the ability of fine-tuned models to represent European weather regimes at sub-seasonal lead times and assess their potential to improve predictive skill in this challenging forecast range. This work is part of the broader EIC-backed TAILOR project, which aims to develop a Finetuning-As-A-Service (FAAS) platform and we discuss how these advances form a key building block within the broader TAILOR vision.