Data driven topology optimization of AM parts accounting for process-affected fatigue performance: Application to automotive and aerospace components

The present paper presents two novel data-driven topology optimization (TO) procedures to design lighter additively manufactured (AM) fatigue resistant components. The first TO method is driven by a probabilistic machine learning (ML) algorithm based on a Bayesian Neural Network (BNN), trained on fatigue data from the literature to assess probabilistic stress-life (PSN) curves. These curves are used to predict the allowable design stress for TO and are predicted directly from AM process parameters, the risk volume, and thermal and surface treatments. The second TO design procedure is instead driven by another BNN, trained to predict the maximum critical defect size from the process parameters. The TO limit stress is computed from the predicted critical defect and the threshold stress intensity factor K th . After the TO, the critical stress intensity factor K I in the component is computed and compared against K th , to assess the effectiveness of this design procedure. These two frameworks are applied to the design of an SS316L automotive suspension lower control arm and a Ti6Al4V aerospace bracket, respectively. With the following framework, the limit stress calculation does not require specifically designed experimental campaigns and prototyping, as previously sparse experimental knowledge can be embedded in a powerful design tool, which allows for preventing fatigue failures, while accounting directly for the influence of the AM process parameters.