Microsoft open-sources a tool to improve 2 to 6 week weather forecasts

In short: Microsoft published and open-sourced a new tool that uses machine learning to correct errors in existing weather models, with the goal of improving forecasts two to six weeks ahead.

What happened

Microsoft researchers, along with partners including MIT and the European Centre for Medium-Range Weather Forecasts (ECMWF), published results on a method called Probabilistic Bias Correction, or PBC, in the journal Nature. The team also released the software as open-source, meaning anyone can use and adapt the code.

Weather forecasts often get much less reliable beyond about two weeks. This happens with both traditional physics-based models and newer AI models. Both kinds of systems have to fill in missing information, like gaps from places with few weather stations, and those guesses can create a consistent tilt in the results, like a scale that is slightly off.

PBC is designed to spot that kind of built-in tilt, also called bias, and correct it. One example Microsoft gives is that a model trained on older weather records may lean too cold because the planet has warmed, so PBC can help adjust for that.

Microsoft says the tool can run on CPUs, which are the standard processors in most computers, rather than needing specialized chips. In tests, the researchers found PBC improved ECMWF’s leading physics-based model for 98 percent of precipitation targets and 92 percent of temperature targets. The analysis also found it doubled the accuracy of ECMWF’s newer AI forecasting system.

Why it matters

If forecasts are more dependable weeks in advance, people can plan earlier. That includes farmers deciding when to plant and harvest, and emergency teams preparing for storms and floods. The fact that PBC is open-source and can run on more common computers could also make these improvements more realistic for weather agencies with limited budgets, including in developing countries.

Source: The Verge AI