Bayesian optimisation of part orientation in additive manufacturing

Abstract

Additive manufacturing (AM) remains slow in terms of volumetric processing rates. Minimising support for overhanging faces is an effective method of reducing material wastage and post-processing cost. Mindful design can remove much of this support; however, well-selected build orientations are still essential. Searching all feasible orientations is inefficient due to the large number of faces in many mesh files. Nevertheless, support structure generation forms a critical part of the AM process planning stage. This research uses novel combinations of proxy evaluation criteria models for support estimation and optimisation methods to minimise support structure. The number of overhanging facets, overall support structure length and a new estimate for the volume of support structure are used in place of a precise calculation of support quantity. These proxies are used within three different optimisation schemes: grid search, random search and Bayesian optimisations (BO). BO is found to out-perform random and grid search techniques, with grid search having the worst performance in most cases, requiring up to 17-times fewer optimisation iterations. The overall length of support is the most effective proxy model, even outperforming the volume of support estimation, and is shown to perform within 3.5% of results benchmarked against commercial software.