According to current ideas about String Theory, is that the standard model is but one vacuum in millions of the theory.

It strikes me that to find the correct vacuum and search through all the possibilities is something that a neural network would be very good at.

We might not be able to solve it with gradient descent if the vacuums are discrete entities not joined by a continuum. But other methods of training the neural network could be used.

Has anyone tried to use neural networks or genetic algorithms to try to find the correct vacuum for String Theory or is the math just too difficult?

  • 7
    $\begingroup$ "It strikes me that to find the correct vacuum and search through all the possibilities is something that a neural network would be very good at." Why does that strike you? Why, stipulating that one could make the process of evaluating "string vacua" completely algorithmic (which you haven't provided the slightest argument for), would a neural network be better at it than a straightforward algorithm that just iterates over all vacua? $\endgroup$ – ACuriousMind Feb 6 '18 at 18:15
  • 3
    $\begingroup$ Because neural networks are good at this kind of stuff. $\endgroup$ – zooby Feb 6 '18 at 18:21
  • 2
    $\begingroup$ Attempts that I have recently seen, this and this. $\endgroup$ – secavara Feb 6 '18 at 18:37
  • 6
    $\begingroup$ @Steeven That reply is just "because neural networks are magical fairy dust that makes everything work forever" without any substance at all. ACM points out very clear flaws in OP's question (which take it from bad to terrible, really) and OP's reply is just to repeat the party line that machine learning is going to save the world, without even bothering to specify what "this kind of stuff" actually is? Neural networks are good at solving highly structured problems with good training data and clear objective criteria of success; OP has not demonstrated any of those conditions hold. $\endgroup$ – Emilio Pisanty Feb 6 '18 at 18:52
  • 1
    $\begingroup$ They are not magical. But they are good at finding patterns in large data sets. The string theory landscape is potentially a large data set. It has an input (a random vacuum) and a desired output (the standard model). Hence it seems an ideal candidate for deep learning. @secavara I would accept this as the answer to the question. $\endgroup$ – zooby Feb 6 '18 at 20:09

Maybe not a direct answer to your question because I don't think one exists yet. However, there are a set of strong conditions which one can impose on the vacua constructed from string theory and F theory (e.g. the weak gravity conjecture) based on which one may constrain realistic models of string vacua from the swampland. Also see these notes by Vafa: https://arxiv.org/abs/1711.00864.

This is certainly an interesting avenue to venture into i.e. to see if machine learning, big data and AI could help discern the set of string vacua which resemble our de Sitter universe from the swampland. To this end, see this paper: https://arxiv.org/abs/1707.00655 and these workshops: https://web.northeastern.edu/het/string_data/ and https://indico.mpp.mpg.de/event/5578/

It's too early to give a good answer to this question - but there is certainly growing interest and some hope.


Not the answer you're looking for? Browse other questions tagged or ask your own question.