Automatic object reassembly is an important task in archaeology, robotics and medicine. The creation of a tool for finding fragments that complement each other and assembling them into a more or less complete object would facilitate the work of restorers. Recent works focus on predicting the pose of a set of fragments relative to each other or on estimating the matching of different surfaces with mathematical approaches. With this article, we propose a deep-learning method for pairwise matching classification of point clouds. The research methodology is based on the use of Siamese neural networks combined with graph convolutions. Pairs of matching and non-matching point clouds have been created using the fragmented objects provided in the Breaking Bad dataset. These objects are represented as point clouds and are divided into multiple pieces. Our best experiment reveals an accuracy of 82.7% and an F1 score of 81.09%, which demonstrates the power of the implemented method.