A neural network for structural re-analysis

Abstract The need for a computationally efficient method of structural re-analysis is long-standing. In general traditional methods of analysis have not proved suitable for the purpose. Recent progress in neural computing technology has provided a more suitable background from which to develop a general iterative re-analysis method. A simple neural network architecture conveniently accommodates design changes in geometry, element/material properties, topology, applied loading, supports and is therefore suitable for insertion in a combined design/re-analysis computing environment. The weight in the neural network represent structural flexibilities and after training a set of weights corresponding to the initial state of the structure, design changes are made as required the network then continues training from the previous state. Illustrative examples demonstrate that the network approach, whilst considerably slower than band-matrix processing, offers considerable advantage in convenience for the designer.