Borrowing from Astronomy to Improve Deep Learning Weather Prediction
, Professor, University of Washington, Department of Atmospheric Sciences
Deep learning weather prediction (DLWP) is an attractive alternative to numerical weather prediction (NWP), particularly at forecast lead times greater than two weeks. The skill in multi-week and seasonal forecasts comes not from knowing today's weather, but from long-lived forcing at the surface, such as El Niño. Long-range forecasts must also be probabilistic, which is typically achieved by making several similar forecasts, known as an “ensemble,” spanning the range of possible future weather patterns. DLWP’s advantages are that it is orders of magnitude more computationally efficient than NWP, allowing much larger ensembles, and it can directly incorporate data about the Earth system that's hard to use in conventional NWP. Astronomers have identified a 12-sided representation of the spherical heavens that we have applied to the Earth. This grid has greatly improved the accuracy and efficiency of our DLWP model, allowing our forecasts to approach state-of-the-art NWP.