Development of Machine Learning Algorithms for Estimating Failures Due to Complex Loading or Geometry

Machine components usually operate under complex states of stress/strain. This is mainly because of two reasons: having complex geometry or being subjected to complex loading conditions. Geometrical discontinuities due to the presence of notches weaken structures. These notches create triaxial stress state that may influence the fracture behaviour. On the other hand, application of multiaxial loading cyclic loading can lead to significant reduction in fatigue life. For instance, 90 out-of-phase loading is commonly regarded as the most detrimental multiaxial loading path. Research efforts have been focused on these two critical problems. As for fracture of notches, there are three major methods that are widely used to estimate the quasi-static fracture load for notched geometries: the strain energy density (SED), the theory of critical distance (TCD) and the notch-stress intensity factor (NSIF). These theories require stress analysis that are usually performed using finite element analysis. On the other hand, low cycle multiaxial fatigue damage parameters have been developed based on strain or energy to analyse multiaxial fatigue behaviour. Strain-based damage parameters such as Fatemi-Socei or Smith-Watson-Topper are called critical plane models because they are evaluated at specific planes. These models have been shown to be successful not only in estimating multiaxial fatigue life but also in predicting fatigue cracking orientation. Such methods require critical plane search technique to identify the plane of maximum normal or shear strains. Similar to conventional models, machine learning approaches such as artificial neural networks (ANN) can also be used to model the material behaviour under quasi-static and complex loading. ANN models are used extensively to model non-linear material responses such as flow, texture evolution, corrosion-fatigue etc. The application of ANN models in various applications is due to the inherent flexibility of these models to model any system. In addition, published works have shown that ANN models provide a faster and more accurate solution when compared to conventional models. As most fatigue models do not show promising results for complex geometry (notched samples) or complex loading, a successful ANN model would reduce experiment costs and provide effective and fast material response predictions. The objective of this research proposal is to develop machine learning algorithm capable of estimating quasi-static fracture load of notch specimens and fatigue life of specimens subjected to strain-controlled multiaxial proportional and non-proportional loading. Quasi-static fracture experiment on a selected strategic alloy will be performed. Notched specimens will be machined with different sizes and geometries. Full strain-field will be measured for selected notched specimens using digital image correlation system. The algorithm will be trained using available and collected experimental data from the literature. Standard material characterisation including grain size, stress-strain curve, and fracture surface analysis will be performed using optical and scanning electron microscopes. Selected specimens will be examined using X-Ray computed tomography (CT) to understand the fracture behaviour. The merit of the developed machine learning algorithms will be examined by comparing their results with that obtained from conventional estimation methods and that obtained from the experiments.