Game physics engines update their objects (acceleration, collision, constraints/joints etc) only so often at a certain tick rate to leave enough cpu time for other things. You can have very precise and realistic physics simulations already - they just don't run at realtime.

I don't think this is a solution for bad physics in games.

The idea here isn't making better collision simulations. We do that well enough already. I think the big deal here is teaching an AI how to predict the outcomes of collisions by training it with a physics engine.