Enhancing Information Flow in Graph Neural Networks for Scientific Machine Learning
M. Chenaud, J. Alves, Frédéric Magoulès · Civil-comp conferences · 2025
The propagation of information through discretised domains is of crucial importance in the field of scientific machine learning.Recent studies have demonstrated the efficacy of graph-based models for physical simulations, particularly due to the inductive biases inherent in such frameworks.However, ensuring efficient information flow through these graph architectures is a delicate aspect, due to the wide range of scales of the simulated phenomena.We summarise some key architectural choices that are the most prominent in the literature, and we propose a novel edge augmentation technique, based on farthest point sampling and the Möller-Trumbore algorithm, for highly nonconvex geometries.The efficacy of our approach is demonstrated through the training of a graph neural network model on a challenging, non-convex geometry.