Today predictions of the next several days’ weather can be remarkably accurate, thanks to decades of development of equations that closely capture atmospheric processes. However, they are not perfect. Data-driven approaches that use machine learning and other artificial intelligence tools to learn from past weather patterns might provide even better forecasts, with lower computing costs.
Although there has been progress in developing machine learning approaches for weather forecasting, an easy method for comparing these approaches has been lacking. Now Rasp et al. present WeatherBench, a new data resource meant to serve as the first standard benchmark for making such comparisons. WeatherBench provides larger volume, diversity, and resolution of data than have been used in previous models. The researchers hope that WeatherBench will also foster competition and collaboration and spur advances in the field.
These data are pulled from global weather estimates and observations captured over the past 40 years. The researchers have processed these data with an eye toward making them convenient for use in training, validating, and testing machine learning–based weather models. They have also proposed a standard metric for WeatherBench users to compare the accuracy of different models.
To encourage progress, the researchers challenge users of WeatherBench to accurately predict worldwide atmospheric pressure and temperature 3 and 5 days into the future—similar to tasks performed by traditional, equation-based forecasting models. WeatherBench data, code, and guides are publicly available online.
The researchers hope that WeatherBench will foster competition, collaboration, and advances in the field and that it will enable other scientists to create data-driven approaches that can supplement traditional approaches while also using computing power more efficiently. (Journal of Advances in Modeling Earth Systems (JAMES), https://doi.org/10.1029/2020MS002203, 2020)
—Sarah Stanley, Science Writer