Modélisation prédictive des défauts géométriques en fabrication additive : application aux structures lattices

Additive manufacturing (AM) offers new possibilities for producing parts with architectural geometries, such as lattice structures. However, controlling the geometric quality of the resulting parts remains a major challenge due to the inherent defects of the manufacturing process. Lattice structures, in particular, have additional geometric defects related to their architecture. The main objective of this PhD thesis is to characterize the geometric defects of lattice parts produced by metallic AM using a predictive model. First, the different geometric defects, their characteristics, and the methods for acquisition of their data by dimensional measurement are presented. Digital processing of the data obtained by measurement or simulation involves a number of operations, in particular the registration process which has been investigated, with the development of a specific method adapted to lattice structures. In addition, a skeletonization technique has been used to improve both registration and partitioning. Geometric deviations are then evaluated and used to train a 2D convolutional neural network. The 3D geometric data are transformed into 2D images via a flattening step and an augmentation technique is applied to enrich the dataset. Furthermore, an inpainting method has been used to refine the data and ease the learning process. Finally, a 3D convolutional neural network was implemented to better adapt to the three-dimensional nature of the geometric data.