AI is changing how quickly weather forecasts can be produced. Recent machine-learning systems have achieved accuracy comparable to conventional forecasting methods in reported tests, while generating predictions in seconds instead of hours. Their promise is significant, but questions remain about how they will perform when the weather departs from familiar patterns.
Forecasts in seconds
Traditional weather forecasting relies on large, complex computer algorithms built around atmospheric physics. Running those models takes hours. AI forecasting systems use a different approach: they learn from historical weather data and can generate a forecast in seconds.
Three systems have brought renewed attention to this approach: Huawei’s Pangu-Weather, Nvidia’s FourcastNet, and Google DeepMind’s GraphCast. Pangu-Weather can also forecast the path of tropical cyclones, extending its use beyond general weather predictions.
The European Centre for Medium-Range Weather Forecasts (ECMWF) model is considered the gold standard for forecasts up to 15 days ahead. Pangu-Weather achieved comparable accuracy to that model. Google DeepMind said in a non-peer-reviewed paper that GraphCast beat it 90% of the time in the combinations tested. These results have led meteorologists to reconsider how machine learning might fit into forecasting.
Speed does not settle the question
A faster forecast can help make the work of prediction less time-consuming. But speed alone does not establish how reliable a system will be across every kind of weather. The AI models described here learn from weather records stretching back decades, so they are well suited to patterns resembling what appears in that history.
That strength also points to an uncertainty. As climate change makes conditions increasingly unpredictable, future weather may not closely resemble the past data used to train the models. Peter Dueben, head of Earth system modeling at ECMWF, said it is not known whether AI models will be able to predict rare and extreme events.
That leaves a central test for AI forecasting: whether its advantages hold when events are unusual, not just when conditions look familiar. The source does not establish that these systems can handle such cases. For now, their strong results in comparisons do not resolve that open question.
A role alongside conventional models
Dueben sees a possible way forward in combining AI tools with traditional weather forecasting models. The two approaches could be used alongside each other to help produce the most accurate predictions. This would let forecasters explore the speed of machine learning while retaining the established physics-based process.
The prospect is relevant because weather affects more than the daily forecast. Oliver Fuhrer, head of the numerical prediction department at MeteoSwiss, the Swiss Federal Office of Meteorology and Climatology, points to economies’ growing dependence on weather. Renewable energy is one reason; logistics and even the number of searches for ice cream are other examples he gives of businesses connected to weather.
AI could help speed up a painstaking forecasting process, particularly when used with human expertise. That combination matters: the models offer rapid calculations based on past observations, while experts remain part of how forecasts are used and interpreted. The source describes this as a potential benefit, not a guarantee that AI will replace conventional forecasting.
Public data and an open future
Weather records are collected and kept by countries, leaving plenty of publicly available data for training AI models. That data gives researchers and technology companies a basis for developing systems such as Pangu-Weather, FourcastNet, and GraphCast.
What comes next is not yet clear. The source points to both scientific exploration and possible services or business models as reasons companies are entering the field. For forecasting, the practical question is how machine learning can contribute alongside existing methods, especially when the weather is difficult to predict from historical patterns.
AI has shown that it can produce forecasts quickly and perform strongly in the comparisons reported so far. Whether it can help reliably anticipate rare extremes remains unresolved. The next stage will depend on how these tools work with traditional models and human expertise as forecasters assess their capabilities.