Scientists use machine learning to speed up weather models, not replace them

In short: Weather and climate scientists are using machine learning to make some modeling tasks faster, but they still rely on physics and human checks.

What's going on

Machine learning, which is a way for computers to learn patterns from past data (like learning what a “typical” storm looks like), is becoming more common in weather forecasting. These systems are not the same as chatbots that write text. They are trained on past weather snapshots to predict the next one.

One example is the European Centre for Medium-Range Weather Forecasts, which put a machine-learning forecast model into daily use in February 2025. It runs alongside the center’s older physics-based model. ECMWF has said a run of its traditional model can use about 1,000 times more energy than its machine-learning model, and it can take about 30 minutes instead of about three.

But there are clear limits. A machine-learning model does not “understand” physics rules, so it can produce impossible results, such as negative rainfall. Researchers add guardrails, for example forcing negative rain predictions back to zero. Another concern is extreme weather, since rare events may not show up enough in the training data, and studies have found some machine-learning models can underestimate record-breaking events.

For climate science, the situation is tougher. Climate models ask “what if” questions about a future we have not seen before, so past data alone is not enough. Teams like Caltech’s CliMA are using machine learning only for specific sub-tasks inside larger physics-based climate models, like modeling snow, while keeping physics rules in place.

What to watch

Expect more hybrid approaches, where machine learning acts like a fast assistant for certain steps, and physics-based models provide the backbone. The key test will be whether these tools can handle extremes and remain trustworthy when conditions change.

Source: Arstechnica