PSI - Issue 60

7

S. Mahesh et al. / Procedia Structural Integrity 60 (2024) 382–389 Mahesh et al. / Structural Integrity Procedia 00 (2019) 000 – 000

388

As discussed in the previous sections the idea of this work is to develop a proof-of-concept where, the capability of ML to predict the FCGR curve is explored. Therefore, limited data is considered for training and validation. The prediction results are shown in Fig. 4a and Fig. 4b for C and m values, respectively. The predicted results are compared with the experimental results of Malipatil et al. (2021). It is interesting to note that predicted C values exhibit wide scatter for GTM 718 while for GTM 720, the scatter is almost nil. This shift in the prediction might be occurring due to ‘ overfitting ’ caused due to very small dataset. This can be avoided by using a larger dataset for training. However, the overall prediction seems to follow a trend similar to that of the experimental data for GTM 718. From Fig. 4 (a) and (b), we see that there is a unique trend exhibited by both alloys GTM 720 and GTM 718 for C and m, respectively. But it can be easily pointed out that there is no correlation in the trend followed by GTM 720 to that of the trend followed by GTM 718, the reason for this is to be looked from microstructural perspective to make any sense. Otherwise, this is a statistical data that is used to train the algorithm. Since in the present work the focus was to train the ML algorithm to achieve a prediction no specific importance to the trend was given. More experimental data is needed to establish a generalized and accurate prediction. Unlike Fig. 4b which shows excellent correlation between be the experimental and predicted m -values for both the materials, Fig .4a is fraught with scatter in the predicted C-values for GTM 718 material. It is evident from Fig. 4 that predicted C values (in Fig. 4a) are independent of the predicted m-values as the latter seems to follow a trend without much scatter. Further investigation is necessary in this regard. As the current study is focused mainly on the prediction of Paris regime of the FCGR curve, the constants C and m were only considered rather than da/dN and ∆K parameters. da/dN and ∆K could also be used for prediction, but even a small error in the prediction would directly be indicated by a visible shift in the FCGR curve. Whereas, prediction error in C and m values does not contribute to huge shift in the Paris regime as seen in the current study. Based on the predicted C and m values, the Paris region of the FCGR plots were constructed and compared with the experimental plots as shown in Fig. 5a and Fig. 5b for GTM 720 and 718 alloys, respectively. The predicted Paris region clearly shows effect of R value of the crack growth rate (da/dN) and is comparable with the experimental results (Boyce and Ritchie (2001)). This indicates that the BPNN model could be used to predict the Paris regime of the fatigue crack growth rate curve and does not seem to be affected by the scatter in the predicted C and m values (Fig. 5). For any further changes in input variables this ML algorithm could provide a consistent prediction instead of performing the experiments. This enables us to reduce the number of experiments that is needed to be performed in order to generate an FCGR curve for each condition, thus reducing the cost of material and experimentation required. Now that the proof-of-concept is developed and ML prediction is proven to be quite successful for FCGR, the study will be extended to a larger set of data and materials. Influential parameters such as material composition, processing condition, microstructure parameters, testing conditions etc. can be utilized to train the algorithm in a more exhaustive manner. Further work is needed to explore the prediction efficacy for regions 1 and 3 of the FCGR plot. 4. Conclusion The current work uses back propagation neural network (BPNN) algorithm to train and predict the Paris constants, C and m of the fatigue crack growth rate (FCGR) curve for two nickel based super alloys viz., GTM 720 and GTM 718. Based on the predicted C and m values, the Paris region was reconstructed. The following are the conclusions from the present work: • It is found that with limited training data, the ML model was able to predict the C and m values with good accuracy and comparable with the experimental findings available in the literature. • The reconstructed Paris region of the FCGR curve clearly shows the effect of stress ratio which is widely accepted for such material in the literature.