Here is the poster that our team-member Loh Sher-En Jessica will present at ICML 2024 : ICML poster

GNNs are one of the state-of-the-art surrogates for numerical CFD simulations. In this work, we propose two novel geometric representations, Shortest Vector (SV) and Directional Integrated Distance (DID), that provide a global geometry perspective to the nodes in GNNs. We also introduce Finite Volume Features (FVF) in the graph convolutions to enable message-passing operations to adjust to different nodes.

Experiments show that the proposed techniques help boost the performance of SOTA GNN-based methods by up to 41%.