PSI - Issue 80

Vinit Vijay Deshpande et al. / Procedia Structural Integrity 80 (2026) 327–338

335

9

Vinit V. Deshpande et al./ Structural Integrity Procedia 00 (2019) 000 – 000

Fig.8. The process flow of a conditional GAN model implemented to determine current density in the composite specimen. During the training process, an input image is fed to the generator that produces a fake output ̂ . The discriminator takes this fake output ̂ and compares it with real output (calculated using FEM) to check how close the fake output is to the real one. The Loss function of the entire model is, ℒ= ℒ ( , ) + ℒ 1 ( ) (12) where, ℒ ( , )= , [log ( , )] + , [log(1− ( ( , )))] , (13) ℒ 1 ( )= , , [‖ − ( , )‖ 1 ] (14) Here, is the expectation operator and ‖ ‖ 1 is L1 norm. In this work, cGAN model is used to predict the current density of the 2D composite specimen studied here. The cGAN model was trained on 2000 images of different composite specimens. Fig. 9 shows the prediction of the current density for 5 composite specimens (a-e) as calculated from FEM simulations (named as ‘ Ground Truth ’) and the prediction of cGAN (named as ‘Prediction’ in Fig.9) and a difference image that shows the difference in the current density values from both methods at each pixel position. It can be seen in the difference images that most of color contour in the particles is blue color. This indicates that the cGAN model was able to accurately replicate the current density contours as seen in the ‘ Ground Truth ’ images. Further, for the same specimens, Resistor Network models were created. Table. 2 shows the resistance of the 5 composite specimens (Fig.9a-e) as calculated from FEM results, the Resistor Network and the cGAN models. In order to calculate resistance from the FEM and cGAN models, current density was integrated along an equipotential line to obtain the total current flowing through the specimen. Then from the current value, the resistance was calculated from the potential difference. It can be seen that the % relative error for the cGAN model is much smaller than the Resistor Network model.