Has anyone tried to train a neural network to learn physics?

[BWV]

Have been curious about a thought experiment where, given enough experimental data to train, a sophisticated enough neural network / deep learning program could 'discover' most of classical and quantum physics. Any thoughts?

 

[newjerseyrunner]

That'd be particularly difficult because to train a neural network, you have to be able to at least in some way quantify how correct or wrong the output currently is, or else it can't do back propagation. It'd be very difficult to describe in any meaningful way how "wrong" a result might be.

You have to be able to at least start the NN off with something that it's trying to do. They'll likely be used in the future to aid with experiments and do trial and error at speeds that no human could even hope to, ala using lasers to create a Bose-Einstein condensate. A neural network replicated a nobel prize winning experiment, did it better than the humans did, and took only an hour.

 

[BWV]

You have the answers in the data, the NN would train against whatever parameter you want it to solve for, then it would just be finding the quantitative relations. Basic mechanics would be fairly trivial take F=MA, train it on a dataset of three vectors M,A,F where F is the Y variable then feed it new M,A data and it will get the relation

But given that NNs universal approximators of nonlinear functions, I would think, given enough data, you could train on some complex physics, say weather or fluid dynamics.​