Generative AI Could Predict Weather a Month Ahead
Researchers at the AIRI Institute of Artificial Intelligence have developed a neural network called Marchuk that can forecast weather anomalies up to a month in advance, and it can run on a standard home laptop.
With its vast territory stretching from the Baltic to the Arctic and a wide range of climate zones, Russia has a strong need for reliable forecasting systems. This new development is designed to support more precise scenario-based weather analysis.
Marchuk works on principles similar to generative video models. The system analyzes four “snapshots” of atmospheric conditions over the past 24 hours, taken at six-hour intervals, and produces forecasts for one to eight days ahead. Its key advantage, however, lies in estimating the probability of climate anomalies over a horizon of up to 30 days.
The “AI meteorologist” has a compact architecture with 276 million parameters. Despite its relatively small size, it matches and in some cases outperforms LaDCast, a much larger model with 1.6 billion parameters.
Such climate modeling could help emergency services prepare for events like flooding by providing detailed scenario forecasts. Energy providers could better assess wind loads, while farmers could take early action to protect crops from drought. The model is named after mathematician Gury Marchuk, a pioneer of numerical atmospheric modeling.
Forecasting for Everyone
According to a VTsIOM survey, 88% of Russians check weather forecasts regularly, and 84% of those users consider them accurate. Researchers are actively working to improve both reliability and long-term forecasting, and there have already been tangible advances.
For example, researchers at the Artificial Intelligence Institute of the Moscow Institute of Physics and Technology have introduced a climate analysis platform capable of predicting extreme weather events such as hurricanes and heavy rainfall with highly localized precision, down to individual buildings. The system identifies relationships between weather patterns and environmental factors such as high-rise buildings, asphalt coverage, and green spaces, then refines broader forecasts into street-level predictions.
Scientists from Southern Federal University and Saint Petersburg State Marine Technical University have demonstrated, using data from 1961 to 2023, that neural networks can effectively handle long-term temperature forecasting tasks.
In parallel, researchers at the Institute of Natural and Technical Systems in Sevastopol have developed Russia’s first neural network model capable of generating detailed forecasts of precipitation, temperature, and flood risks up to nine months ahead. The model analyzes 70 years of climate data and updates its forecasts on a monthly basis.
A Question of Safety
Deploying such neural network models is directly tied to saving lives and reducing economic losses. In recent years, AI-driven meteorology has advanced rapidly worldwide. NVIDIA has introduced FourCastNet, capable of producing weekly forecasts in seconds. Google DeepMind and Huawei have developed GraphCast and Pangu-Weather, which are tens of thousands of times faster than traditional numerical models.
The European Centre for Medium-Range Weather Forecasts (ECMWF) is advancing hybrid forecasting approaches, widely seen as the future of meteorology. However, many countries lack access to expensive supercomputing infrastructure. Marchuk, which can run on a standard laptop, could offer a practical export solution for regions that need to deploy forecasting systems quickly without major capital investment.
AIRI’s model targets the challenging sub-seasonal window of up to 30 days. In practice, Marchuk is designed to complement traditional numerical models used by Roshydromet, enabling faster scenario analysis.
