Development of Machine Learning Models for Prediction of Flaw Interaction

Abstract In the case of detection of flaws in pressure components, flaw characterization is crucial for assessing flaw acceptability. When detected flaws are in close proximity, Fitness-for-Service (FFS) Codes provide proximity rules to determine whether the flaws should be combined or considered independently. Most proximity rules rely on a limited number of parameters, such as flaw depths and the relative distance between flaws, to keep the criteria simple. Recently, ASME Code Case N-877-1 was developed to include the flaw aspect ratio as an additional parameter influencing flaw interaction. This Code Case is based on a mechanical interaction threshold (a 6% increase of Stress Intensity Factor) derived from experiments. However, the proximity rules of Code Case N-877-1 still include some conservatism inherent to their formulation, which must remain easily applicable. Since mechanical interaction between flaws is usually calculated based on three-dimensional Finite Element analyses, surrogate models may be useful to instantly predict flaw interaction level for any combination of influencing parameters, aiming to broadly quantify the conservatism of proximity rules. The object of this paper is to develop such surrogate models using machine learning approaches. The input data set used to train and test the models comes from eXtended Finite Element Method (XFEM) analyses of flaw interaction. While a comprehensive set of flaw configurations was used to develop Code Case N-877-1, additional calculations are considered to extend the surrogate models’ scope and enhance their efficiency. Three machine learning models are compared. Their accuracy, determined from a test data set, is assessed to identify the best surrogate model for predicting flaw interaction.