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 equati

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.