تقسيم بندي حلقه هاي بدون ساختار براي مدلهاي اجزاء محدود با استفاده از شبكه هاي عصبي مصنوعي

Partitioning of Unstructured Finite Element Meshes Using Artificial Neural Networks

In this paper the use of neural networks for partitioning of adaptive, unstructured finite element meshes for parallel time-stepping finite element analysis is presented. The concept of mean filed annealing and its use in finding approximate solutions to combinatorial optimization problems is investigated. The application of mean field annealing neural network for the partitioning of finite element meshes is described. This partitioning is based on the recursive bisection approach. The mapping of the mesh bisection problem onto a recurrent neural network is presented. The use of a trained backpropagation neural network which can estimate the number of elements which will be produced inside each coarse element of a finite element mesh after mesh refinement is described. The solution quality and the computational time of solutions are demonstrated using a case study.