> I accept (I think) your point that ML and simulation can be used to "discover" those laws in some inductive sense. Yes, we use these methods to confirm that

> I accept (I think) your point that ML and simulation can be used to "discover" those laws in some inductive sense. Yes, we use these methods to confirm that reality matches some candidate law. I personally don't think it's very meaningful to talk about physical laws existing in some independent platonic sense, but the debate over the principle of induction and the epistemology of science is a side story here IMHO. Ultimately what we care about is that a model successfully predicts reality in the future. I think a useful conceptual division is: with some models, the internal structure of the model gives us understanding of the physical behavior. Most of what we call "laws of physics" are in this category. We can look at the algebraic expression and understand symmetries, constants, conserved quantities, etc. A second type of model is purely algorithmic and utilitarian: it gets the right answer, but the internal structure of the model doesn't provide any comprehensible information or understanding. Most ML algorithms fall in the latter camp. What muddies the issue is that experimental physics tries to validate models of the first kind, but now generally requires methods of the second kind. We're forced into these methods because of the complexity of what we're studying (particle physics can involve thousands of Feynman diagrams), the noisiness of our measuring equipment, and the rareness of the events we're searching for. When I say physics methods are very relevant to wall street, I'm talking about the latter methods more than any attempt to find fundamental theorems.