A comparative study of machine learning methods for predicting mode I and II brittle fracture in notched bend specimens.

Abstract

This study evaluates the efficacy of statistical, artificial neural network (ANN), and deep learning (DL) methods in developing mathematical models for estimating dimensionless fracture parameters in single-edge notched bend (SENB) specimens under Mode I/II brittle fracture conditions. The research focuses on this specimen, a commonly used configuration in fracture mechanics studies. The primary objective is to benchmark the accuracy, computational efficiency, and generalization capabilities of various modelling approaches, including linear regression, polynomial regression, Random Forest Regression (RFR), and Bidirectional Long Short-Term Memory (BiLSTM) networks. The study leverages a comprehensive dataset encompassing various geometric configurations and loading conditions to develop predictive models for key fracture parameters such as Y_I, Y_II, and T^*. The results shows that regression methods are insufficient for capturing the complex, non-linear relationships inherent in fracture mechanics. RFR demonstrated superior accuracy with training R² values of 0.99 (Y_I) and 0.99 (Y_II), and validation R² values of 0.93(Y_I), 0.96 (Y_II), and 0.99 (T^*). BiLSTM also performed robustly, achieving validation R² values of 0.99 (Y_I), 0.96 (Y_II), and 0.99 (T^*). In contrast, simple regression methods like multiple linear regression (MLR) and polynomial regression (PR) showed limited effectiveness, with R² values as low as 0.44 (MLR) and 0.57 (PR) for Y_I.