@WebHubTel, I think there is a fourth possibility to those the geologist suggested:

4) Model the real world using one or more closed-form solutions which have no (known) physical connections to the real world yet predicts observations well.

5) Model the real world as a tree (perhaps but not necessarily a random tree or random forest) with closed-form solutions applicable to tiny bits of physics at each of its leaves. In practice, there would be a large library of tiny-bit-of-physics models, and the trees would be grown based upon data using boosting, and possibly model averaging.