Neural Network Systems to Reproduce Crash Behavior of Structural Components

The use of neural networks as global approximation tool in crashworthiness problems is here investigated. Neural networks are not only asked to return some meaningful indices of the structural behavior but also to reproduce load–time curves during crash phenomena. To contain the number of examples required for the training process, parallel subsystems of small neural networks are designed. Design points for the training process are obtained by explicit finite element analyses performed by PAMCRASH. The settlement of the points in the design domain is defined using a maximum distance concept. The procedure is applied to different typical absorption structures made of aluminum alloy: riveted tubes, honeycomb structures, longitudinal keel beam and intersection elements of helicopter subfloors.