they do but ..
most of what modern AI is good at is modeling things that repeat on a regular basis.
it’s a bit of a parlor trick
they have a ground truth
they then train these systems to match the ground truth
and all the wiggle room around the ground truth captures a variety of human variety.
so things like faces are mostly predictable. they have ranges.
very simple Physics is very predictable so they can imitate universes.
flow is far too complicated yet.
they have laminar flow OK.
but turbulence is one of the greatest mysteries in the sciences and they had a lot of trouble with it.
it’s very difficult to model stirring creamer into coffee
they do experiments with modeling folding different colored clay in the 1970s
after just a few folds they start losing track
they’ve gotten much better in the past 50 years and they can model hurricanes surprisingly well considering their complexity
but they’re basically stirring creamer in a coffee cup
clouds and things are much more diverse