PSI - Issue 75

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

516

8

layers are used for the final prediction. The optimiser used for this model is 'adam' and the chosen cost function is 'mean_squared_error' (Fig. 8.)

Fig. 8. AI model architecture diagram

Out of the 1000 data files, 500 are used to train the AI model. The objective is to enable the model to predict the yield strength and the tangent modulus of the bilinear constitutive laws that were used to generate the strain fields. In this way, the model learns to determine the mechanical properties based on the strain fields. Once the training is completed, the remaining 500 files are used to test the model’s performance. In this stage, only the strain fields are provided, and the model is tasked with identifying the corresponding constitutive laws. The predicted results are then compared with the actual values used to generate these fields in order to evaluate the accuracy of the model. With this method, the cost function converges rapidly towards zero, and satisfactory model performance is achieved after a limited number of 'epochs'. This model is adapted for each part of the specimen (HAZ, weld bead and base metal) to allow separate training and evaluation for each zone. The experimental strain fields are then processed to match the format of the numerical, normalised fields. Once the AI model is trained, the experimental fields can then be used as input data for it, in order to determine the actual constitutive laws for the three parts of the welded zone. This method allows the evaluation of three local bilinear elasto-plastic constitutive laws, which corresponds to 6 physical quantities: 3 yield strengths and 3 tangent moduli. These predictions are consistent with the experimental results from image correlation; indeed, a significantly lower yield strength is found in the weld bead, which is consistent with the strain concentration observed in this zone. The three constitutive laws highlight a decrease in yield strength in the fusion zone compared to the base metal, and conversely, an increase in the tangent modulus. Indeed, the yield strength difference is noted as 9.5% between the base metal and the HAZ, and 15% between the base metal and the weld bead. For the tangent modulus, its increase is 18% between the base metal and the HAZ, and 31.7% between the base metal and the weld bead, respectively. 3. Conclusion The results of the preliminary measurements have justified the importance of taking into account local constitutive laws. The analysis of the experimental stereo digital image correlation results has allowed the qualitative highlighting of differences in local mechanical properties at the welds, notably through the study of strain localisation. Finally, the use of a numerical model coupled with AI allows the quantitative evaluation of local constitutive laws at the welds.