Simulating Structural Analysis with Neural Network

Abstract

Structural analysis programs used in solving design problems are often computationally expensive. Obtaining optimal solutions typically requires numerous iterations involving analysis and optimization programs. This process becomes prohibitive due to the amount of computer time required for convergence to an optimum design. Any new techniques significantly reducing the computer time required to solve design problems would be beneficial. One promising technique is to simulate a slow, expensive structural analysis program with a fast, inexpensive neural network. Guidelines for designing and training a neural network to simulate a structural analysis program are developed. These guidelines include the selection of training pairs and determining the number of nodes on the hidden layer. A sample problem shows that by following these guidelines, a neural network can reduce the amount of time it takes an optimization process to converge to an optimum design.