Jakub Lengiewicz

Abstract: Deep learning approaches are being increasingly used for accelerating scientific simulations, however, they fail to perform efficiently as the size and complexity of the problem increases. In this talk, we introduce a novel geometric deep-learning framework for performing supervised learning on graph-structured data. Our approach is based on the Multi-Channel Aggregation (MAg) operation, which efficiently handles non-linear regression tasks on graph-structured data. The MAg operation is combined with our novel clique-based graph pooling layers to create a graph U-Net, named MAgNET, which is both robust and able to handle arbitrary complex meshes. Importantly, MAgNET scales efficiently with problem size. We evaluate the performance of MAgNET against Perceiver IO, a state-of-the-art attention-based neural network architecture from the same family of architectures as ChatGPT.

Saurabh Deshpande is a doctoral researcher at the Department of Engineering in the Faculty of Science, Technology and Medicine of the University of Luxembourg. His research is focused on the development of novel scalable deep learning techniques for their applications in computational mechanics and computer vision.