Convolutional Neural Networks for Part Orientation in Additive Manufacturing

The industrial applications of additive manufacturing (AM) have seen noticeable growth in recent years, pushing the studies on the parameters affecting the aesthetic, functional, and structural characteristics of the produced component. A central role is attributed to the orientation of the geometry on the building platform and the resulting building direction, the choice of which usually relies on the expertise of the operator. This work aims to elaborate an algorithm able to predict autonomously the optimal positioning of the component through the innovative techniques of deep learning, employed for their ability to draw information from a set of examples and build complex models. A convolutional neural network (CNN) is developed that, starting from the tridimensional representation of an object, predicts the rotation angle pair that leads to the optimal printing configuration. Two approaches have been compared: the first one predicts the angles, represented as points on a unit circle, through a regression that minimizes the angular difference loss. The second one performs a classification over a set of discreet rotations. The algorithm is trained and validated on two different datasets; finally, the generalization capacity of the model is investigated, highlighting the limits linked to the choice of examples used during training.