Supervised Machine Learning Approaches for Structural Integrity: Residual Stress Evaluation and Defect-based Fatigue Modelling

Assessing the performance of structural materials is imperative to guarantee the integrity, service continuity, durability and load-bearing capacity of engineered parts. Amongst numerous failure mechanisms, fatigue is reportedly one of the most detrimental and catastrophic since it gradually and relentlessly damages structural components until they suddenly collapse. The fatigue performance evaluation is rendered convoluted by numerous concomitant impacting factors. In this respect, the scientific community has widely recognised the critical role of residual stress (RS) and manufacturing defects, which alter the local stress state induced by externally applied cyclic loads and lead to scattered fatigue response. Therefore, the key objective here is to route the development of a comprehensive fatigue design setting with the aim of including RS and defects in future engineering practice. Hereby, diverse supervised Machine Learning (ML) techniques are harnessed while prioritising probabilistic aspects to enhance the robustness of the proposed approaches. To incorporate RS, users must ascertain the trustworthiness of their evaluations. However, this stage is frequently hampered by deterministic and heavily user-dependent data regularisation protocols applied during the pre-processing of input data. To address this issue, Gaussian Process Regression is exploited as a stochastic regularisation technique. To showcase its potency, this technique is applied to the Contour Method, whose potential and cost-effectiveness are incredibly attractive. Although entrenched Continuum Solid Mechanics theories can capture the influence of defects on fatigue, they often restrict the number of explanatory defect descriptors, usually to a characteristic length. Moreover, the related models hinder an exhaustive quantification of the involved uncertainties. Therefore, traditional fatigue models are revisited and opportunely treated through the lens of ML. Specifically, defect-based approaches to estimating the finite fatigue life and fatigue endurance of metallic materials are conceived and appropriately addressed through diverse variants of Neural Networks, and Logistic Regression. The present work, therefore, provides the research & engineering community with systematic ML-supported predictive tools, thus constituting a step towards an integrated framework for probabilistic design against fatigue.