Advancing the analysis of architectural fabric structures, neural networks and uncertainty

Abstract

In current practice a plane stress framework comprising elastic moduli and Poisson’s ratios is most commonly used to represent the mechanical properties of architectural fabrics. This is often done to enable structural analysis utilising commercially available, non-specialist, finite element packages. Plane stress material models endeavour to fit a flat plane to the highly non-linear stress strain response surface of architectural fabric. Neural networks have been identified as a possible alternative to plane stress material models. Through a process of training they are capable of capturing the relationship between experimental input and output data. With the addition of historical inputs and internal variables it has been demonstrated that neural network models are capable of representing complex history dependant behaviour. The resulting neural network architectural fabric material models have been implemented within custom large strain finite element code. The finite element code, capable of using either a neural network or plane stress material model, utilises a dynamic relaxation solution algorithm and includes geodesic string control for soap film form-finding. Analytical FORM reliability analysis using implied stiffness matrices' derived from the equations of the neural network model has also been investigated.