Forecasting the Weather with AI Analysis of the European Weather Forecast Model
Current weather predictions solve fluid dynamic models of the atmosphere and water vapor and clouds and the earth’s topography on supercomputers to predict the weather up to 10 days in advance. The first computers were even applied to this in WWII. But for a long time, experienced weathermen could take the atmospheric picture and apply their knowledge to make fair weather predictions. That job is now being done by applying Artificial Intelligence learning to 40 years of data and model analyses by the European Center for Medium-Range Weather Forecasts (ECMWF).
Once trained, the AI can be run on a desktop in a minute to give forecasts 6 hours ahead with at least 90% of the accuracy of the European Model, and sometimes even beating it. However, the one million grid points covering the earth amount to cells of 28 km x 28 km at the equator, or 17 x 17 miles. This can skimp on a lot of important geographical features, such as shorelines or mountains.
Apps are already available on laptops, but charge fees. These can even predict the weather that you will encounter on a trip.
Google DeepMind has a graphical machine learning application called GraphCast. This was even better at predicting tropical cyclones, atmospheric rivers, and extreme temperatures. These have to be done rapidly and locally to warn of rapidly changing dangers.
GraphCast can predict 6 atmospheric variables at 37 levels of altitude, including temperature, wind speed and direction, and humidity.
Here is a European area weather prediction for the atmospheric pressure from GraphCast in contours, and the wind speed in m/s in colors at altitudes where the pressure is 850 hPa (1,000 hPa being close to sea level.)
As usual with AI, there is not an understanding of what the AI neural network has really learned. This may be a problem in using these tools as the climate changes, which it already seems to be rapidly doing.
