The race is on to bring AI to weather forecasting
Driving the news: The U.K. Met Office, which already runs one of the top weather forecast models in the world, announced a new partnership on Tuesday with the Alan Turing Institute to develop highly accurate, lower cost forecast models using machine learning and AI techniques.
- Rather than being run on supercomputers and based on the physics of how the atmosphere works, AI-driven weather models are instead trained on historical data and don't need to be run on supercomputers.
- They use current data, algorithms and other techniques to then determine what weather systems are likely to do.
- In the U.S., NOAA is also examining how its forecasters can utilize AI, and it is finding novel, low-hanging fruit applications such as translating its forecasts into multiple languages.
Between the lines: Companies such as Google, Microsoft, Nvidia and Huawei have shown progress in advancing weather forecasting through AI during the past two years, claiming their models meet or beat the top-rated model out of the European Center for Medium-Range Weather Forecasts (ECMWF).
- The ECMWF has gone so far as to make public AI-generated forecasts that are fed with the center's initial data. Working with universities, NOAA is likely to do so soon, except using its Global Forecast System model, the Washington Post reported.
- The Met Office as well as NOAA are not seeking to upend their supercomputer-based numerical models anytime soon. Rather, they are more likely to first develop AI-based models to augment their forecasts for specific purposes, such as high-resolution short-term forecasting, where the larger-scale models have struggled.
- In addition, they can evaluate AI-generated models for their accuracy.
- Already, forecasting successes, such as correctly showing the path of Hurricane Lee well in advance of government-run models, have caught the forecasting community's attention.
The intrigue: The key advantage that AI-driven models offer is to make obsolete the main limitations involved with numerical weather modeling, which is the computing power and time they require. It can take mere minutes to generate an AI-powered forecast, compared to hours of supercomputer time for numerical models.
Yes, but: There are risks that AI-driven advances are being oversold. In fact, they could prove susceptible to errors that current models are not.
- Skeptics of AI weather and climate models say they are essentially black boxes, with the output only as good as the data set they were trained on.
- There is also the concern that such models could fail to recognize anomalous events because of biases or gaps contained in the data they learned from.
The bottom line: The vast savings in computing power and therefore expenses that AI-generated forecasts make it extremely appealing to weather and climate agencies trying to make forecasts more accurate.