Predicting climate conditions months in advance remains one of the most difficult problems in Earth science. A recent study, published in npj Climate and Atmospheric Science proposes a new approach, combining artificial intelligence with probabilistic modelling to improve seasonal forecasts.
Traditional methods rely on general circulation models (GCMs), which simulate atmospheric and ocean dynamics. While physically grounded, these models are computationally expensive, limiting their scalability. Simpler statistical models are more efficient but often less reliable due to limited observational data. This trade-off has long constrained forecasting performance.
A hybrid, data-driven approach
The new framework, developed in the context of the AI4DROUGHT project, takes a different route by learning directly from data. It combines observational records with outputs from existing climate simulations, effectively expanding the training dataset. The model is built on transformer-based neural networks and uses variational inference to represent uncertainty.
Instead of producing a single forecast, it generates probabilistic predictions, capturing a range of possible outcomes and their likelihoods. This allows the system to reflect the inherent uncertainty of climate processes rather than masking it.
Palma, L., Peraza, A., Civantos-Prieto, D. et al. Data-driven seasonal climate predictions via variational inference and transformers. npj Clim Atmos Sci 9, 48 (2026). https://doi.org/10.1038/s41612-026-01320-z
Featured image : ACC (Anomaly Correlation Coefficient) against ERA5 reanalysis (2001–2021), for temperature (2T, top row) and precipitation rate (PR, bottom row). Seasons are shown in the columns: DJF (December–January–February), MAM (March–April–May), JJA (June–July–August), and SON (September–October–November). Black dots indicate statistical significance at the 95% confidence level. Credit: https://doi.org/10.1038/s41612-026-01320-z