Researchers have applied transformer neural networks—the architecture behind ChatGPT and similar AI systems—to seasonal climate forecasting. The new approach, developed under the European Space Agency’s AI4DROUGHT project, predicts temperature and precipitation anomalies up to three months ahead.
The study, published in npj Climate and Atmospheric Science, represents the first use of vision transformers for seasonal prediction. Unlike weather forecasting (days ahead), seasonal prediction tries to identify whether the coming months will be warmer or wetter than normal—information valuable for agriculture, water management, and disaster preparedness.
How It Works
Traditional seasonal forecasts rely on general circulation models (GCMs)—physics-based simulations of the atmosphere and oceans. These are computationally expensive and struggle with some regional predictions.
The new framework combines variational inference, a statistical technique for representing uncertainty, with transformer architectures originally designed for processing images. The model was trained on output from CMIP6 climate simulations, then validated against ERA5 reanalysis data—the best available reconstruction of historical weather.
One key innovation: the system generates probabilistic predictions showing a range of possible outcomes rather than single-point forecasts. This better reflects the inherent uncertainty in predicting climate months ahead.
The approach also explicitly separates the contribution of long-term climate change trends from natural variability. This matters because a model that simply predicts “warmer than average” will be right most of the time due to global warming, without providing useful information about specific seasonal anomalies.
Where It Works
The model was tested globally across all four seasons. For temperature predictions, much of the skill comes from capturing climate trends—not surprising, given that warming is the dominant signal in recent decades.
Precipitation forecasts showed more nuanced results. The AI system provided “added value in numerous extratropical inland regions” where traditional teleconnection patterns (like El Niño’s influence) are weaker. These are areas where conventional forecasting systems often struggle.
However, the AI model underperformed SEAS5, the European operational forecasting system, in tropical regions. The researchers position their approach as complementary to existing systems rather than a replacement.
The team also applied the method specifically to Europe, comparing performance at different spatial resolutions.
What This Means
Seasonal climate prediction has long been a weak link in the forecasting chain. Weather models are good out to about a week. Climate projections handle decades to centuries. But the months-ahead timeframe—when farmers plant crops, utilities plan energy demand, and emergency managers prepare for drought or flood—remains difficult.
AI approaches trained on climate model output could help fill this gap, particularly in regions poorly served by current systems. The probabilistic nature of the predictions also gives users a more honest picture of forecast uncertainty.
The Fine Print
Several limitations temper enthusiasm. Temperature skill is heavily influenced by the climate change signal—the model’s predictions are less useful for identifying year-to-year variability. In tropical regions, traditional physics-based systems still outperform the AI approach.
The model was trained on simulated data from climate models, not direct observations. While this expands the available training dataset, it also means the AI inherits any biases in those underlying simulations.
Perhaps most importantly, seasonal prediction remains fundamentally hard. Even with AI, forecasting whether July will be wet or dry in a specific region involves large uncertainties that no model can eliminate.
The research was funded through the European Space Agency’s AI4DROUGHT project.