For more than half a century, weather forecasting has been an exercise in applied physics. Meteorologists built enormous computational models that simulate the atmosphere by solving equations of fluid dynamics, thermodynamics, and radiative transfer across millions of grid points. These numerical weather prediction systems—run by agencies like the European Centre for Medium-Range Weather Forecasts and the U.S. National Oceanic and Atmospheric Administration—represent one of humanity's great scientific achievements. They are also, increasingly, being outperformed by neural networks that have never solved a differential equation in their lives.
The shift began in earnest when Google DeepMind's GraphCast and Huawei's Pangu-Weather demonstrated that AI models trained purely on decades of historical weather data could match or exceed the accuracy of physics-based systems for forecasts up to ten days out. More remarkably, they did so in minutes rather than hours, running on a fraction of the computing power. What once required supercomputers costing hundreds of millions of dollars now runs on hardware a well-funded startup could afford.
The pattern-matching paradigm
Traditional forecasting works from first principles: given the current state of the atmosphere, calculate what physics dictates will happen next. AI forecasting works differently. It learns statistical relationships between past atmospheric states and their subsequent evolution, then applies those patterns to new observations. The model doesn't know why cold fronts behave as they do; it simply recognizes the signature of conditions that historically preceded certain outcomes.
This distinction matters philosophically but perhaps less practically than purists might hope. Weather systems, for all their complexity, exhibit regularities that machine learning excels at detecting. The atmosphere has been doing roughly the same things for billions of years, and a few decades of comprehensive satellite and sensor data capture enough of those patterns to be predictively useful.
Where physics still wins
AI forecasting has clear limitations. Models trained on historical data struggle with genuinely unprecedented conditions—the kind of extreme events that climate change is making more frequent. They also lack the interpretability of physics-based systems; when an AI model makes a poor forecast, diagnosing why is difficult. Traditional models, by contrast, can be interrogated step by step.
The hybrid approach emerging at major forecasting centers uses AI for rapid initial predictions while retaining physics-based models for extreme scenarios and for understanding the mechanisms behind forecasts. The European Centre has begun integrating machine learning into its operational pipeline, treating AI not as a replacement but as a complementary tool that excels in different conditions.
Our take
The weather forecasting revolution offers a useful template for understanding AI's broader impact on scientific domains. Machine learning doesn't replace domain expertise; it forces experts to articulate what their expertise actually consists of. Meteorologists are discovering that much of what they do involves pattern recognition that machines can automate, while the genuinely irreplaceable parts—understanding causation, handling novel situations, communicating uncertainty to the public—become more visible and more valued. The forecast for AI in science is similar: disruption first, then a clearer division of labor between human and machine intelligence.
