The project Replica, about six years ago, paved the way for computational studies of visual patterns in art history. Simultaneously, it created the possibility for art historians to trace the propagation of patterns throughout the history of art. During the project, an image retrieval network was set up to discover artistic patterns given an input image. Despite successfully serving monographic needs and targeted search attempts, the network does not propose spontaneous discoveries. In this thesis, we eliminate the middleman of the input image, creating clusters of artworks sharing a common pattern propagation. The clusters are integrated further with a 2D coordinate-based visualisation, which provides an organic view of the evolution of the patterns in art history. In this effort, we demonstrate the effectiveness of fine-tuning deep learning models on a set of visual connections using a compound Hinge loss and ResNeXt architecture. Moreover, we show that clustering the trained visual signatures with OPTICS yields remarkable precision. We emphasise the importance of the semi-supervised learning of the clusters, proving the qualitative and quantitative improvement over generic clustering methods. Furthermore, we close the loop of the semi-supervised clustering through the annotation of the new findings in the clusters proposed, and retraining thereof. In total, we add over 700 new images to the set of slightly over 1800 existing visual connections. We find, in addition, examples of cross-domain, architectural, design and sketch-based patterns, which were previously outside the scope of the known visual connections.