PSI - Issue 57

Mohammad F. Tamimi et al. / Procedia Structural Integrity 57 (2024) 121–132 127 Mohammad F. Tamimi & Mohamed Soliman/ Structural Integrity Procedia 00 (2023) 000 – 000 7

A MATLAB script (MathWork, 2020) is employed to build and execute the FE model and to capture the output (the crack driving force in terms of the J -integral) in response to the input parameters. Convergence analysis indicates that a feedforward ANN with 8 hidden layers and 7 neurons per layer accurately represents the relationship between input parameters and FE-derived crack driving parameters. The trained ANN demonstrates satisfactory accuracy in predicting the crack driving parameter, with an average error below 5% as indicated in Figure 5(b) when 45,000 samples are used. Figure 5(b) presents the regression plot, emphasizing the close alignment between the network output and the FE prediction, highlighted by the a strong fit along the 45° line. Table 1. Range of input parameters covered in the ANN training dataset.

Range [150−450] mm [5−30] mm [5−30] mm [0−450] mm [50 − 450]mm [5−30] mm [37.5 − 830] mm [0−100] MPa

Input Parameter Stiffener Spacing

Stiffener Web Thickness Stiffener Flange Thickness Stiffener Flange Width Stiffener Web Height Main Panel Thickness Crack Size (Main Panel)

Stress Level

Fig . 4. A schematic representation of the general layout of multilayer feedforward ANN.

Fig. 5. : (a) Relationship between error percent and the number of samples in the trained dataset; (b) Comparison between the FE results and ANN prediction analyses based on the training data.