Local hemodynamics is mechanistically linked to pathological mechanisms underlying coronary artery disease. Within this context, wall shear stress (WSS) features have been associated with the onset and progression of cardiovascular diseases. WSS is typically quantitatively assessed trough computational fluid dynamics (CFD) simulations in 3D models of coronary arteries reconstructed from clinical images. However, numerical simulations are computationally time-consuming, thus often incompatible with the clinical practice. Therefore, the clinical application of hemodynamic parameters such as WSS is still hampered by the required computational time. In this context, deep learning (DL) approaches have been recently proposed as promising strategies to estimate WSS replacing CFD simulations. Here, the employed dataset is composed of 1078 patient-specific single-branch, diseased coronary arteries reconstructed from coronary angiography. Transient CFD simulations were performed to compute WSS using a finite element-based code (CAAS Workstation WSS software, Pie Medical Imaging) and finally time-average WSS along the cardiac cycle was computed. The aim is to extend previous investigations on patient-specific coronary arteries exploiting a DL-based framework to obtain coronary WSS estimation in “real-time”, with the addition of physical information into the data-driven framework. More in detail, a geometric DL approach will be adopted, based on graph neural networks to learn the relationship between shape and WSS. We aim to perform: (1) hyperparameters optimization, (2) benchmarking and scalability tests on HPC systems, and (3) evaluation of the performance of the DL-based framework on image-based coronary models. The following submission is a unique collaboration between KI and KTH proposing a new research project with a dedicated student now involved. As such, this proposal is independent on each separate allocations and and research lines, therefore we request a separate allocation for this specific project.