General Circulation Models (GCMs) are expensive - it would be great if we could forecast and simulate the weather with cheap machine learning models instead.
So, take a batch of reanalysis output, and train a neural net to estimate the weather 6-hours ahead from the weather now - to make a forecast (source: github.com/philip-brohan/weather2weather).
That gives us a 6-hour forecast method, and by cycling the model output as the input for the next iteration, we can forecast the weather as far into the future as we like.
The video shows temperature anomaly (warm=red, cold=blue), mean-sea-level pressure (contours) and precipitation (green) from the output of such a ML system.
Does it work? Well, no, but maybe if we tried a little harder?