PSI - Issue 60

Balaji Srinivasan et al. / Procedia Structural Integrity 60 (2024) 418–432 Balaji Srinivasan et al. / Structural Integrity Procedia 00 (2019) 000 – 000

428

11

(Beyond STP Limits)

Maximum Outer Surface Stress (MPa) - BRANCH

Maximum Outer Surface Stress (MPa) - BRANCH

100 150 200 250

1000 1200

0 200 400 600 800

0 50

6

7

11

1 2 3 4 5 8 9 10

STP FEA Pred

FEA Pred

Fig. 11. STP vs FEA vs Predicted Maximum Outer Surface – Header (Within STP Limits) and FEA vs Predicted Maximum Outer Surface (Beyond STP Limits)

4.2. Discussion: From the investigation it is found that XGBoost regression and Random Forest regression models predict the stress factors with minimum error for STP stress factors data. For building a complete prediction model for total 30 stress factors, a best possible combination of XGB and RFR models are used, and further predictions has been done. From Figures 6-11, the ML predicted values are almost in the range of the STP stress factors. For local membrane stress prediction of header, ML predicted stresses are almost lower than STP predicted values and above actual FEA predicted values. For the dimensions beyond the STP limits, the ML predictions are in close agreement with the FEA results. The behavior is same for the local membrane stresses at branch as well. For prediction of maximum inner and outer surface stress factors for branch and header beyond STP limits are conservatively higher than the FEA results. However, for Model #7, where both D/T and t/T are beyond the STP limits, the predicted model computed unconservative results than the actual FEA results. Still, the predicted values are within the 10 % of accuracy of FEA results. Using machine learning technique, the predicted values are within data sets of STP model. Beyond the STP limit, the predictor model with the available data, the predictions are almost inline with the FEA computed results which cannot be computed from using STP. For Model #6, d m /D m is 0.8 which is beyond the limit of ≤ 0.7 as per Eq. (2), and Model #11, t/T is 11, which is beyond t/T limits of 10, the ML predicted values are inline with FEA results. For Model #7, D m /T and t/T are beyond the limits as per Eq. (1) & Eq. (3), where the predicted results are not conservative as compared to actual FEA computed results. 5. Conclusion This paper showcases the successful application of machine learning techniques to address real-world engineering challenges. The machine learning models effectively predict local stresses in nozzles, shells, and formed heads based on external loads, utilizing computed finite element output data sets from FEA (STP). The study also examines various manual analytical techniques, their applicability, and limitations. The comparative analysis highlights the superiority of advanced regression models, XGBoost and Random Forest, in stress prediction. Additionally, the ML model's extension beyond the limitations of the STP data sets proves its versatility and practicality. By leveraging FEA results-based data sets, this research opens exciting possibilities for stress prediction in various components without complex Finite Element Analysis for each instance. This advancement holds significant potential for reducing reengineering time and computational costs using machine learning models.