This is fascinating, thank you.
So is it the case that the AI system would look at a given weather situation and all the inputs this entails, and then construct a probability-based forecast? (So to put it very crudely, if input 'x' has, 9 times out of 10, produced output 'y,' the prediction will therefore be 'y'. Obviously the real model would have to balance huge quantities of such outputs.)
And is AI really purely based on historic data? Surely a system could also improve on its own accuracy by comparing its earlier forecasts to the actual weather that then occurred. So real time weather data could inform two sets of modelling, retrospective and predictive...