Probabilistic Machine Learning for preventing fatigue failures in Additively Manufactured SS316L

• The R90C90 BNN separates the experimental uncertainty from the ML model uncertainty. • The SLM process parameters influence on the life is effectively forecasted by the BNN. • The BNN enhances the cyclic response predictions by generalizing well the ML model. • The PINN allows to improve fatigue life prediction over a FFNN. This study presents a probabilistic machine learning approach to predict and improve the fatigue performance of additively manufactured SS316L components. By analyzing key manufacturing parameters as process settings, thermal treatments and surface treatments, the developed models provide statistical estimations of fatigue strength, that provides valuable insights to extend fatigue life. Trained on an experimental database of fatigue tests, the Bayesian Neural Network (BNN) is employed to separate model uncertainty, related to data limitations, from the uncertainty inherent in the fatigue phenomenon. This approach robustly predicts Probabilistic Stress-Life (PSN) curves, offering valuable insights into the impact of manufacturing parameters on fatigue resistance, allowing to further postpone fatigue failures. The results demonstrate increased robustness and trustworthiness compared to other deterministic machine learning models, making this method suitable for critical applications where failure prevention is crucial.