4-5-1 Neural Network results
4-5-1 Neural Network results
- The second NN configuration had structure 4-3-1:
- input layer with 4 nodes for ε, ε’, T and curve peak strain εp
- hidden layer had 5 nodes
- output layer had 1 node for flow stress s prediction
Desired and predicted flow stress vs. strain for test 28
To consider the influence of characteristic experimental curve features on the material behaviour modelling, curve peak strain εp was added to the input features and 4-component input vectors ε,ε', T, εp were used for training and testing the 4-5-1 NN. The εp value used in the leave-k-out NN training and testing was obtained by averaging the p values of the training curves, i.e. all curves but the one left out for testing. This procedure was adopted because the p values for the available training set did not show any dependence on temperature or strain-rate in the considered variation ranges. Desired flow stress and predicted flow stress were plotted vs. strain. Curve reconstruction is still poor in curve geometry (the reconstructed curve is still a straight line) but the predicted flow stress is very close to the actual one, at least at large strains. Also the 4-5-1 NN is unable to predict the work hardening and work softening material behaviour. However the addition of p seems to provide the NN model with information useful to reduce the offset between predicted and desired flow stress, at least in the steady state region
Who's online
There are currently 0 users and 150 guests online.