Re: Is ML the answer?
No, it is not.
ML is perhaps a very nice approach for weather but is bad for climate. Reason for this is two things (1) purpose of weather prediction is to predict weather, not to be science (no-one cares if prediction system uses tealeaves), and (2) training data is endless – you predict weather n days out and in n days you have training data to improve that prediction. And because weather prediction is short timescale you can assume say composition of atmosphere is constant on those timescales.
Climate is different. First of all training data is less easy to come by: if you predict climate in 2046 under assumptions of atmospheric composition changes xyz over time and other changes tuv over time, then to get training data for this both takes a very long time, may be undesirable experiment to do, and is mutually exclusive with any other training data unless you can clone universe, which is expensive. You can check model against past climate and of course people do this.
And also climate models are (must be) science. If your model predicts x then it is important that you can go into model and say, OK, model is solving these equations using this solver and this data and this is why x is prediction. And if you think x is unrealistic you can then say, OK what is wrong with our solver, is our cloud model inadequate (yes it always is) and so on and so on. If all you can say is 'there is this opaque blob of trained weights in some neural network' that is very unsatisfying.
Biting the hand that feeds IT
