CERN Summer Student Project Report - Albert Sund Aillet

This summer student project consisted of three subprojects. First, two introductory subprojects were conducted as an introduction to the Tensorflow environment, the graph generation, graph batching and the training loop. Then followed a subproject that aimed to model a physics simulation of connected vessels. Finally an image processing subproject was conducted where the goal was to detect cracks from shortest-paths tree graphs generated from images of road cracks. In all of these subprojects, a so-called EncodeProcessDecode model from the Graph Nets Library was used with different numbers of processing steps. The first of the two introductory subprojects showed that the Graph Neural Network (GNN) could correctly sort the values represented by a fully connected graph. The mistakes made by the GNN occurred when two of the node values were close to each other and they were being predicted as being connected when they should not. The second introductory subproject showed that the GNN could learn to highlight the shortest path in a graph with weighted edges, but as in the first introductory subproject, the GNN sometimes highlighted two paths with similar lengths. The physics simulation subproject was performed by multiple experiments concluding that a GNN can reliably approximate a simulation of a physical system without any prior knowledge of it, simply by being trained on observations of the system. The crack detection subproject was performed by different runs of input data and target data of images of road cracks and resulted in the finding that it is possible for the GNN to highlight the underlying cracks in the shortest-tree graph of parts of the image. The summer student project was successful in showing that GNN methods have potential to be of use in a wide range of problems and can be useful for various tasks at CERN.