MAE Department - PG Seminar - Scalable, adaptive, and explainable scientific machine learning with applications to computational fluid dynamics
Zoom Link: https://hkust.zoom.us/j/95465784424?pwd=VEhRbU82VUlIZVp1YXlqNjNJc0cxUT09
Meeting ID: 954 6578 4424
Passcode: 069088
Scientific machine learning (SciML) involves developing physics-informed data-driven algorithms to enhance computational workflows. This talk presents recent results addressing three critical limitations of state-of-the-art SciML algorithms: scalability to realistic scientific computing problems, learning from unstructured mesh-based data, and interpretability. To tackle these challenges, we propose a novel multiscale graph neural network that approximates functions on unstructured and potentially adaptive meshes with high degrees of freedom, providing interpretability by identifying crucial subgraphs for predictions. Additionally, we take concrete steps towards a-posteriori model error estimation by linking these subgraphs to areas contributing to significant spatiotemporal testing errors. The significance of overcoming these limitations is also emphasized, as we strive to revolutionize computational approaches in scientific domains.
Romit Maulik is an Assistant Professor of Information Science and Technology with an appointment in the Institute for Computational and Data Sciences at the Pennsylvania State University. In addition, he is also a Joint-Appointment Faculty at the Mathematics and Computer Sciences Division at Argonne National Laboratory. His research group, the Interdisciplinary Scientific Computing Laboratory, is primarily focused on solving large-scale computational science problems using data-intensive scientific machine learning.