Title:Stress, Strain, or Displacement? A Novel Machine Learning Based Framework to Predict Mixed Mode I/II Fracture Toughness

Abstract:Accurate prediction of fracture toughness under complex loading conditions, like mixed mode I/II, is essential for reliable failure assessment. This paper aims to develop a machine learning framework for predicting fracture toughness and crack initiation angles by directly utilizing stress, strain, or displacement distributions represented by selected nodes as input features. Validation is conducted using experimental data across various mode mixities and specimen geometries for brittle materials. Among stress, strain, and displacement fields, it is shown that the stress-based features, when paired with Multilayer Perceptron models, achieve high predictive accuracy with R2 scores exceeding 0.86 for fracture load predictions and 0.94 for angle predictions. A comparison with the Theory of Critical Distances (Generalized Maximum Tangential Stress) demonstrates the high accuracy of the framework. Furthermore, the impact of input parameter selections is studied, and it is demonstrated that advanced feature selection algorithms enable the framework to handle different ranges and densities of the representing field. The framework's performance was further validated for datasets with a limited number of data points and restricted mode mixities, where it maintained high accuracy. The proposed framework is computationally efficient and practical, and it operates without any supplementary post-processing steps, such as stress intensity factor calculations.