A Hybrid Modelling Tool for Prediction and Managing Distortion in Metal Additive Manufacturing

Metal Additive Manufacturing (AM) has become a key technology in manufacturing of parts or components having complex geometries. In this process, uncontrolled distortion can cause deviation from dimensional tolerances of built parts and may lead to higher rejection rates. This becomes more significant in large-format metal AM.

Accurate predictions of residual stress and distortion in additively manufactured parts and components are usually based on fully coupled physics-based thermal and structural Finite Element Analyses (FEA) of the AM processes. However, the computational demands of high-fidelity simulations can be prohibitive, especially during the iterative design and optimisation stage.

Recent advances in deep learning provide a promising alternative through the training of surrogate models to approximate FEA responses, thus bypassing numerical solution methods. In this paper, we propose a hybrid approach, in which physics-based thermal FEA are coupled with surrogate models of structural FEA to predict residual stress and distortion in additively manufactured parts. We utilise the flexibilities of Graph Neural Networks (GNNs) and employ the Graph Sample and Aggregate (GraphSAGE) architecture to build the surrogate models of structural FEA, enabling rapid predictions of residual stress distribution in the additively manufactured parts.

We apply the developed hybrid approach to predict residual stress and distortion in additively manufactured Ti-6Al-4V parts and compare them against those predicted using fully coupled physics-based thermal and structural FEA. The results indicate that a great saving in computational effort can be achieved while the accuracy of residual stress and distortion predictions is dependent on the workflow from data synthesis to model training.