Microsoft tool sharpens forecasts two to six weeks ahead
Microsoft researchers have developed a machine learning tool designed to improve weather forecasts two to six weeks ahead, when existing physics-based and AI systems often become less reliable. The approach could help farmers and emergency responders plan further in advance, while giving weather agencies a less resource-intensive way to improve forecasts.
Called Probabilistic Bias Correction, or PBC, the open-source tool identifies and corrects systematic errors in existing models, including biases caused by relying on historical climate data. In the study, it improved a leading European physics-based model for 98 per cent of precipitation targets and 92 per cent of temperature targets, and doubled the accuracy of the European Centre for Medium-Range Weather Forecasts’ new AI system. Unlike tools requiring supercomputers or powerful GPUs, PBC can run on CPUs, though putting it into routine forecasting remains a challenge.
- Microsoft’s PBC tool aims to improve forecasts two to six weeks ahead.
- It can run on CPUs, making it more accessible to resource-limited agencies.
- The study found gains in precipitation and temperature accuracy.
New here? Start with this
Weather forecasts are produced by computer models that use physical rules and historical climate data to estimate what may happen. Their reliability tends to fall several weeks ahead, making it harder to plan for conditions such as heavy rain or unusual temperatures.
Microsoft researchers developed Probabilistic Bias Correction, an open-source tool that looks for recurring errors in existing forecasts and adjusts for them. It has been tested with a leading European physics-based model and a newer AI forecasting system; it can run on ordinary processors rather than requiring specialist computing equipment.
More reliable forecasts weeks ahead could help farmers plan work and emergency responders prepare for weather risks. Weather agencies would still need to work out how to fit the tool into routine forecasting.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Supporters can argue that better forecasts several weeks ahead would give farmers, emergency responders and public agencies more time to prepare, and that PBC offers a relatively affordable way to improve existing systems. Its reported gains across many precipitation and temperature targets, and its open-source availability, suggest it could make useful forecasting improvements accessible without major new computing infrastructure.
The case against
A cautious case is that results from a study do not by themselves show how well PBC will perform in routine forecasts or across different regions and changing climate conditions. Because it corrects models using learned patterns, there is a risk those corrections may be less dependable when conditions depart from the historical data; agencies would need to validate its forecasts and address the practical challenges of operational use.
AI Business Companies Research Science Technology
Read the full article at the source →
Originally published by The Verge as “Microsoft has a new machine learning tool to improve weather forecasts”.