Neural networks in Wolfram Mathematica

[Vrbic]

TL;DR Summary

Is it reasonable to do specific neural networks in WM?

Hello,
I have a question about the use of functions and overall creation of neural networks in the Wolfram Mathematica (WM) program. I wonder if it is realistic to make meaningful neural networks usable at least partially for scientific research in WM? By scientific research, I do not mean the study of neural networks directly, but their application to some specific mathematical or physical problems. Something like using NDSolve to solve the differential equations of e.g. Einstein's relativity, which may in turn lead to a new discovery or in general to the preparation of a scientific paper.

Or is it just a toy or a teaching aid for very specific tasks such as converting handwritten digits to digital digits shown in the WM tutorial?

Thank you for your opinions.

 

[DrClaude]

I don't know how efficient Mathematica is with neural networks. But as for physics, people are working on trying to get AI to help solve physics problems:
Physics-informed Machine Learning
Physics-Informed Machine Learning: A Survey on Problems, Methods and Applications