Fatigue property prediction of additively manufactured Ti-6Al-4V using probabilistic physics-guided learning

The probabilistic fatigue properties of additively manufactured (AM) Ti-6Al-4V using selective laser melted (SLM) process is analyzed considering the effects of process parameters. The Probabilistic Physics-guided Neural Network (PPgNN) is proposed for the modeling. With this developed model, both mean and variance of the fatigue life can be learned. The PPgNN contains constraints on model parameters to obtain the probabilistic stress-life relationships (P-S-N curves) with the physics-consistent curvature and nonconstant variance. The PPgNN model is also able to be trained using the data set with missing data for more reliable predictions. Experimental fatigue data are collected from extensive literatures for AM Ti-6Al-4V in as-built and annealed condition subjected to various process parameters (scanning speed, laser power, hatch space, layer thickness, heat temperature, heat time). The PPgNN model is validated using the experimental data. Next, a group of models with the same architecture and training data but different initial neural network biases and weights are obtained to account for the the randomness of NN. Following this, the predictive performance is compared between models training using all data (both complete and incomplete) and only complete data. Finally, global sensitivity analysis using the Delta Moment-Independent Measure is conducted to investigate the importance of process parameters. By only varying one process parameter and keeping the remaining ones fixed, the effect of each parameter on probabilistic fatigue lives is studied.