PSI - Issue 75

515 7

Arthur THIBAULT et al. / Procedia Structural Integrity 75 (2025) 509–518 Arthur THIBAULT / Structural Integrity Procedia (2025)

Base Metal

a

Weld

b

HAZ

Fig. 7. Geometry used in the numerical model (a) Numerical Exx strain distribution (b)

Once the simulation is performed (Fig. 7b), strain fields linked to the parameters of the bilinear laws used are obtained. In order to test a large number of different constitutive laws, the combined use of a Python code and an APDL script allows a large number of simulations to be run, changing the parameters of the three constitutive laws each time. For each calculation performed, the displacement and strain fields are extracted In total, 1000 cases were be calculated and served as a database for IA identification; Table 2 shows the bounds used for the constitutive law parameters. These bounds were chosen based on the results of Puydts, Q. (2006) and slightly adapted following the experimental tests. Table 2. Bounds for the behaviours laws parameters used to create the database Area step step

Yield stress min (MPa)

Yield stress max (MPa)

Tangent modulus min (MPa)

Tangent Modulus max (MPa)

Melted area

225 240 250

325 340 350

0.1 0.1 0.1

1500 1500 1000

2500 2500 2000

1 1 1

ZAT

Base metal

1.1.1. Use of the AI model for the prediction of behavior laws After the creation of the database, an AI model is implemented. The aim is to train this model using the database and then apply it to experimental strain fields in order to identify the local mechanical laws in the three zones of the weld. Initially, a machine learning model was considered, but as the results were not satisfying, indeed the prediction of mechanical property wasn’t efficient. Finally, a deep learning AI model was ultimately implemented in order to exploit the previously created database. The AI model used is coded in Python using TensorFlow for the definition of the neural network. It is a model consisting of a Convolutional Neural Network (CNN), which is particularly effective for processing structured data and identifying mechanical properties, as shown by Motamedi et al. (2022). Firstly, this model extracts the numerical results from the previous database for each of the three zones. To be used as input for the CNN model, these data are then organised and normalised. The neural network consists of 4 convolutional layers which apply filters to extract relevant features from the input data. Each of these layers is followed by a pooling layer to reduce dimensionality without loss of information. Next, a layer flattens the data. Finally, 4 dense