PSI - Issue 64

Manish Prasad et al. / Procedia Structural Integrity 64 (2024) 1524–1531 Manish Prasad / Structural Integrity Procedia 00 (2019) 000 – 000

1529

6

to iterate though the data and find out which 80:20 split to choose. K-fold cross validation has been used to perform this task. It splits the data in K equal parts and trains the model with each fold iteratively and the left data to test the model. In this way the fold that produces a better ML performance is established. Training data is used to train each of the four ML regression algorithms discussed before. Each model has been previously checked for overfitting and underfitting. Overfitting is obtained when an ML model performs better on training data than test data, while underfitting is obtained if the model does not reach a minimum level of accuracy. If the algorithm did not fit well, cross-validation step was repeated with a different ML regression algorithm. When the proper fit was achieved (i.e., without over/under fitting), the accuracy threshold was checked. To find a better accuracy, hyperparameters were tuned and the algorithm was trained again. This process w as repeated until the accuracy reached an optimum value, for each of the four presented ML regression algorithms. The regression algorithm with maximum accuracy was defined as the final model to predict the CCS failure load. Such a model was obtained for each of the four ML regression algorithms. These models were then ready to predict the CCS separation load for new specimens with different geometrical and mechanical properties.

Fig. 3. Flowchart of ML procedure

4. Results and Discussion Table 2 presents a summary of the accuracy obtained from the different ML models used in the present work. Accuracy is checked by comparison of R-squared score, mean value of P exp / P model and standard deviation. For comparison purposes, the accuracy of the ML models is also compared to that of existing models in the literature, such as GEP (Al-Ghrery et al., 2021), fib Bulletin 90 (fib, 2019) and Smith and Teng (Smith & Teng, 2002b) formulations.

Table 2. Comparison of ML prediction with other methods

Model KNN

R-squared score

Mean of P exp /P model

Standard Deviation

0.975

1.025

0.078