US government researchers are experimenting with using a form of machine-learning to speed up the performance of a major forecasting model.
The researchers from the US Department of Energy’s Argonne National Laboratory have trialled the use of deep neural networks to enhance the capabilities of the Weather Research and Forecasting (WRF) model, a comprehensive meteorological model used extensively in forecasting and weather research.
Currently, the WRF model relies on a method known as parameterization to model the way that different environmental conditions interact to cause weather. But parameterization is computationally expensive since it involves modeling at a scale far greater than the actual phenomena.
Environmental and computer scientists from Argonne are collaborating to use deep neural networks to replace the parameterizations of certain physical schemes in the WRF model. As well as being less expensive, deep neural networks significantly reduce the simulation time, says Argonne environmental scientist Jiali Wang.
“With less-expensive models, we can achieve higher-resolution simulations to predict how short-term and long-term changes in weather patterns affect the local scale,” Wang told Science Daily. “Even down to neighborhoods or specific critical infrastructure.”
The team first focused its research on the planetary boundary layer (PBL), the lowest part of the atmosphere, which extends up only a few hundred meters from Earth’s surface.
The deep neural network was able to successfully simulate wind velocities, temperature and water vapor. The results also showed that neural networks in one location can predict behavior across nearby locations with accuracy rates of more than 90%.
The team’s ultimate goal is to replace all of the expensive parameterizations in the WRF model with deep learning neural networks.