Prediction of fatigue failure in small-scale butt-welded joints with explainable machine learning

Butt-welded joints are common in many industries. The fatigue behavior of such depends on numerous factors, e.g. load level, local weld geometry, or parent material strength. To make things worse, these factors often interact, yet mutual influence can hardly be quantified by multivariate studies, i.e. varying one factor at a time out due to the large number required tests and statistical nature geometry. Consequently, assessment deviates significantly between prediction experimental result. Thus, alternative methods desirable order take various influencing into account. this end, machine learning techniques were used predict failure locations cycles performed small-scale butt-welded joint specimens. In addition accurate predictions, an understanding importance is desired, ranking most important factors; however, capturing several possibly interacting usually requires complex nonlinear models. We gradient boosted trees. Since black box models, SHapley Additive exPlanations (SHAP) framework was explain identify influential features their interactions. Lastly, model explanations linked back domain knowledge.