Data scarcity in fatigue design of AM parts: A transfer learning framework to model the process − structure − property relationship

The paper demonstrates the effectiveness of transfer learning (TL) in improving the fatigue failure prediction of additively manufactured (AMed) metallic alloys. The approach is applied to three commonly used AMed alloys − Ti6Al4V, AlSi10Mg and SS316L − to investigate the potential of multi-material knowledge transfer in improving the predictive accuracy under data-scarcity conditions, which are typical of fatigue testing. The TL framework employs a Bayesian neural network (BNN) pre-trained on data from two alloys and subsequently fine-tuned using data from the third, data-scarce alloy. Starting from the process parameters, risk volume, thermal and surface treatments, and the applied stress, the BNN predicts the fatigue life at a defined reliability (R) and confidence (C) levels. The TL effectiveness is assessed though a cross-material validation on the three alloys, by changing in turn the two alloys of the pre-training and the third as the target material with a truncated training dataset, mimicking realistic experimental scenarios with limited data availability. The model accuracy is evaluated using the R2 score of the predictions on external validation datasets and is compared against the model trained without TL on the same reduced database. The results shows that TL improves both accuracy and generalization capability, benefiting from heterogenous multi-material knowledge, allowing to design against fatigue failures with greater confidence in this data scarce conditions. Moreover, TL enhances the ability to infer the influence of process parameters on fatigue behaviour, even when the available data are limited.