PSI - Issue 42

Chahboub Yassine et al. / Procedia Structural Integrity 42 (2022) 1025–1032 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

1031

7

The time to generate the database was 300 minutes; each simulation took around five minutes. The simulations were done with a different set of GTN parameters based on the data found in table 1. The mesh size in front of the pre-crack tip is 0.125 mm 0.0625 mm, as well as the mesh, is composed of axisymmetric quadratic elements with eight nodes. ANN and the creation of the database As already mentioned, 60 simulations for NT specimens with different GTN parameters were performed to generate the database essential to train the ANN. The trained ANN model for the NT test has 200 input neurons, 75 hidden neurons, and 5 output neurons (200-75-6). The values of the response force F refer to (neurons of the input layer), and the GTN parameters to be identified are represented by the neurons of the output layer (f 0 , f c , f f , f n , S n , and ε n ). We estimated the GTN parameters after training the neural network in a fraction of the time compared to the direct technique, which combines experimental and finite element data. f 0 =0, f c =0.0045, and f f =0.25, f n = 0.05, S n = 0,45 et ε n = 0.2 are the GTN parameters determined by utilizing the ANN. Prediction of Crack propagation for PIPELINE SPECIMEN The main step now is to predict the pipeline's failure based on the GTN parameters found in the last paragraph. The FEM model has 104193 nodes and 248038 elements in total. We performed a FEM simulation using the GTN parameters obtained by the ANN; the results indicate that the simulation curve closely matches the actual curve and that they agree well in Fig. 6.

3000

Experiment

ANN SIMULATION RESULTS

2500

2000

1500

Load [kN]

1000

500

0

0

2

4

6

8

10

12

14

16

18

20

Fig. 6. Prediction failure using ANN

By analyzing the curve above, it is obvious that ANN estimates the correct values of GTN parameters. It is noticeable that the curve found by ANN does not fully fit the experimental data, especially at the end of the curve. This phenomenon is related to the database used to train the network. We cannot control the database used in the neural network, but it is possible to use other geometries as a database, such as CT specimens and SENT specimens. However, this approach will take more time compared to the NT