Data-driven based fracture prediction of notched components

Abstract

A data-driven approach is developed to predict the fracture load of a notched component. To do so, more than 1500 fracture tests (507 unique experimental data points) on mixed-mode I/II loading of notched brittle samples were collected from the literature. After pre-processing the raw data, six features of maximum tangential stress ( σ θ θ max ) , maximum tangential stress angle ( θ σ θ θ max ) , ultimate tensile strength ( σ u ) , fracture toughness ( K I c ) , notch opening angle ( 2 α ) and notch tip radius ( ρ ) were selected by using the neighbourhood component analysis (NCA) technique. To predict the fracture load of various types of notched samples, several machine learning (ML) models were trained using the methods of Gaussian process regression (GPR), decision tree ensemble and artificial neural network (ANN). Then, the Bayesian optimization algorithm was applied to find the optimum hyperparameters for each model. Lastly, the performance of the models in predicting fracture load was evaluated against 124 unseen data points. The results revealed the high potential of data-driven methods for assessing the fracture load of notched brittle components with acceptable precisions of 92%, 89% and 88% accuracy, respectively, for GPR, decision tree ensemble and ANN models. The superior performance of the GPR method can be attributed to its ability to capture complex nonlinear relationships in the data while providing reliable uncertainty estimates. Furthermore, thanks to its interpolation capabilities, GPR is able to seamlessly fill the gaps between data points, resulting in more comprehensive and precise predictions across the entire range of input data. Additionally, the presented models were capable of predicting the fracture load of VO-shaped notched samples with acceptable accuracy, though this type of notch was not used in the model training process. This article is part of the theme issue 'Physics-informed machine learning and its structural integrity applications (Part 2)'.

Explosive utilization efficiency enhancement: An application of machine learning for powder factor prediction using critical rock characteristics

“…In addition, GPR, utilizing its probabilistic framework, offers uncertainty estimates for its predictions and can effectively handle outliers, making it a powerful tool in decision-making processes [ 72 ]. The GPR model (MD3) proposed in this study consists of one dependent variable, namely a powder component, and six independent variables, as shown in Table 2 .…”