Bayesian model updating and model validation for fatigue life prediction of additively manufactured aluminum alloys

Fatigue life prediction for additively manufactured aluminum alloys is challenging due to limited data and significant material scatter. To address this, this study proposes a Bayesian uncertainty quantification framework for probabilistic prediction and model selection. Fatigue experiments were conducted on Laser Powder Bed Fusion (LPBF) AlSi10Mg specimens under two build directions. Bayesian inference was employed to update the parameters of four candidate models, characterizing data scarcity in the form of prediction uncertainty. U-pooling and Model Evidence metrics were then utilized to quantitatively evaluate model performance. The proposed framework effectively generated 95% confidence intervals that encapsulate most of the experimental data. Results indicate that specimens loaded perpendicular to the build direction (H-direction) exhibit superior fatigue resistance compared to those loaded parallel to it (S-direction). Quantitative validation identifies the Smith-Watson-Topper model and the Morrow model as the optimal predictors for the two directions. This work provides a reliable tool for fatigue assessment under data-scarce conditions.