Re:Superconductivity Theory Under Attack
by eigenspace on Monday December 02, 2019 @12:55AM
(#21680156)
A lot of people are trying to explore modern ML methods in physics, but they’ve so far had limited success at least in terms of interesting theoretical results. One glaring problem is that ML methods tend to be black boxes, so even if someone solved all of condensed matter physics with some super neural network, the question would remain “how and why was this able to work?” “What is going on physically?” which is what physicists tend to actually care about more than being able to make a prediction.
Believe me though, people are trying.
On the other hand, there is plenty of work on rule extraction from trained neural networks being done. Wouldn't it be amazing if we extracted actual laws of physics from a network trained by observations?
Re:Superconductivity Theory Under Attack
by a-nikolaev on Monday December 02, 2019 @12:56PM
(#21684844)
Yes, I know that.
What I'm saying is that if an ML model can approximate quickly the solution, then there could be a simple theory expressed in terms of equations, approximating the full problem. I.e. if there is an efficiently computable procedure, then it's a sign that there might be a good simple approximating theory.
You start with 1D problems, move to small 3D volumes, and build from there. ML is good at "interpolating" or guessing what the answer must be from a given input. Maybe, one will need many nested levels of ML models to realistically simulate a solid-state material, but this is not entirely impossible, I think.
If such an ML simulation can be done reasonably efficiently, then likely there could be a theory that can be formulated in terms of equations, approximating the "theory of everything" with sufficient detail.