I know that nowadays there are some physics engines pretty advanced, capable of very complex simulations.
Are we at a point in technology where if, for example, we were to simulate a rock being dropped on the floor from a certain distance, the simulation can calculate the shape and weight of the rock , the air resistance experienced during the fall, the density of the floor where the rock will fall onto, and all the other thousands of factors involved, and from those things “calculate” the sound that the rock will make when hitting the floor, and then reproduce it?
Is there such a thing? Are we there yet? If not, is it something feasible?
What you can do, and is probably the best way to get this to work, is tackle this with machine learning. You will need lots sound samples of rocks, with details of the rock, and feed them to some (probably deep learning) model.
Speech mimicking with AI has shown we are able to mimick voices, so I think a similar approach would work for rocks. Probably need some tuning and a bit different architecture for nice results since the application differs a bit.
It will of course be an approximation, but that is any calculation. Since all models are wrong, but some are less wrong than others.