Multivariate interpolation and machine learning models for extreme defects-based fatigue life prediction of Ti6Al4V specimens fabricated by SLM

The paper proposes a novel methodology for extreme defects-based fatigue life prediction of Ti6Al4V specimens fabricated by selective laser melting (SLM) technique. The introduced framework is represented by a traditional ranking method based on the three highest values of maximum stress intensity factor ($K_{max}$) with related defect parameters (size, distance from the free surface, compactness and sphericity), training set augmentation using variational autoencoder (VAE) and optimized data-driven models. The defects were observed using micro-computed tomography (µ-CT) prior to the fatigue tests. As data-driven methods a multivariate interpolation and machine learning (ML) models were employed and tuned using Bayesian optimization algorithm called tree-structured Parzen estimator (TPE). The proposed methodology was validated on the test set and the highest prediction accuracy was achieved by random forest (RF) model with value of coefficient of determination $R_{\mathrm{test}}^2=0.956$. Additionally, the SHapley Additive exPlanations (SHAP) analysis was conducted to gain a deeper insights of applied data-driven models.