-1

I have some ideas in the optimization of the design stage of an engineering structure (like a wall). Conventional design methods are really time-consuming if I need to run the finite element-based engineering software several times to find the optimum design parameters of the structure. So I thought to build a machine-learning model using the data generated with this engineering software instead of directly using it for the purpose of reducing the run-time.

My structure can be modeled with 9-12 input parameters (such as dimensions, material properties, forces etc.) and 4-5 outputs (such as displacements, internal forces etc.). If I would like to cover all the ranges for input parameters (with relatively moderate sampling intervals) I need to create millions of training data. This requires my algorithm to run for years in my average-performance computer. This is not efficient and logical!

I would like to ask you for any advice about the choice of any ML model to solve such problems. Can I develop an ML model without covering all the input parameters range that may still present acceptible solutions? How effective to solve the problems with ANNs? Or should I give up applying such ML approach here? Do you have any other approach to offer me checking out?

ouflak
  • 2,458
  • 10
  • 44
  • 49
engineer0688
  • 1

1 Answers1

0

Everything old is new again.

This Google search tells me that combining neural networks with finite elements is both old (2nd citation is from 1992) and new (1st citation is from 2021).

I doubt very much that you "need to create millions of training data".

It could be that NN isn't the best solution for this problem.

My employer was using genetic algorithms to optimize jet engine performance back in the early 90s. That problem seems more complex than the structural problem you describe.

I would encourage you not to throw neural networks at the problem without understanding it better.

duffymo
  • 305,152
  • 44
  • 369
  • 561