PSI - Issue 60

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

384

3

ΔK curve generation , the ΔK value is continuously reduced until the crack growth is negligible or completely stopped. T his value of ΔK is defined as threshold stress intensity factor ΔK th . The regime 2 is a stable crack growth regime and covers the major part of the curve. This region is governed by the Paris law and is given by – = (Δ ) (1) where, da / dN is the rate of crack growth, ΔK is the applied stress-intensity factor range, C and m are material constants also called Paris constants, indicating the Y-axis intercept and slope of the Paris region, respectively. In the second stage the ΔK is gradually increased until the failure of the specimen. The measure of stress intensity factor at failure gives the value of critical SIF K c . This data yields a sigmoidal shaped da/dN vs. ΔK curve on a log log scale as shown in Fig. 1. Thus, an FCGR curve for one particular stress ratio is obtained. To generate FCGR curve for another value of stress ratio, the process is repeated. It is evident from the work of Boyce and Ritchie (2001) that the threshold continuously shifts with varying R. Therefore, the process of testing to obtain FCGR curve for various R values is costly and laborious task. Although predictions can be made for intermittent R values from the experimental data using numerical methods such as Newman equation, Forman equation etc. (Forman et al. (2005)), these are mostly curve fitting models and determining the equation parameters for these models are difficult and time consuming. Thus, considering the broad spectrum of applications and advancements happening in the field of ML, it would be worthwhile to explore the possibility of ML application in FCGR prediction. Today's technological advancement is intertwined with the implementation of ML. Various applications such as speech recognition, image recognition, recommender systems, game development, software integration, chat-bots and so on use ML algorithms at different scales, and AI integration for various other applications is steadily rising. Most of these AI tools run on Artificial Neural Networks (ANN). Generally, an ANN is a multi-layered perceptron derived from the architecture of biological brain, with the inspiration of building smart systems. ANN's are used in large number of applications, from simple classification tasks to identify defective objects in a manufacturing firm, to complex creative tasks for image generation and building personal assistants and chatbots. Advancement of ML has led to its applications in various engineering disciplines. Limited literature is available on the application of ML for FCGR predictions. This is mainly due to lack of detailed FCGR data, for example information regarding the material and processing condition of the specimen, which are needed to train the algorithm. Yet, few attempts have been made to apply ML in the prediction of fatigue crack growth.

Fig. 2. Architecture of a typical BPNN