#2nd place - Offset-based Graph Convolution for Mesh Graph

In the ML4CFD Competition, addressing the challenge of irregular data points in ML4CFD simulations, we opted for a Graph Neural Network (GNN) approach. We constructed graphs for each simulation by integrating the original mesh with a k-Nearest Neighbors (KNN) algorithm. Our GNN architecture comprises a standard 2-layer graph convolution combined with a Multilayer Perceptron (MLP). The innovation lies in our graph convolution method, which aims to emulate 2D convolution while accommodating irregular data points. We introduced an offset-based graph convolution that employs an MLP to generate kernel weights based on the offsets of neighboring nodes, followed by a weighted averaging for aggregation. This method yielded low errors and substantial speedups over traditional PDE solvers, demonstrating the potential of GNNs in enhancing computational fluid dynamics simulations.