AI is increasingly used in weather and climate modeling, but the technology is not as revolutionary as some might suggest. Instead, it relies on established machine learning techniques that have been studied for years. These methods are applied differently in weather and climate simulations, with the latter focusing on long-term trends rather than short-term forecasts. The use of AI in these fields is still in its early stages, with major forecast centers developing their own models to complement traditional approaches. Source: arstechnica
Machine learning models in weather forecasting are trained on two sets of weather data obtained at different times. These models run much faster than traditional weather models because they do not solve complex physics equations at every location. Companies like Google, Nvidia, and Microsoft have developed initial models, often in collaboration with academics. The European Centre for Medium-Range Weather Forecasts (ECMWF) launched its first machine-learning-based model, AIFS, in February 2025, alongside its traditional Integrated Forecasting System (IFS) model. Source: arstechnica
The ECMWF model uses reanalysis data, which creates a physically consistent picture of the atmosphere by combining all available weather observations. This data simplifies the task of predicting future weather conditions based on past patterns. However, machine learning models can produce nonsensical results, such as negative precipitation values, because they do not inherently understand physical laws. To address this, the ECMWF model remaps negative precipitation values to zero. These models also significantly improve computational efficiency, with the AIFS model using about 1,000 times less energy and requiring only 30 minutes compared to the IFS model’s three hours. Source: arstechnica
