GNNs are well suited for challenging problems involving intricate graph structures, such as those encountered in physics, biology, and social networks. By leveraging the structure of graphs, GNNs are capable of learning and making predictions based on the relationships among nodes in a graph. MeshGraphNet architecture based on the work by Tobias et al, The pretrained model checkpoint comes from the 2D Vortex shedding example as described in this example.
The model was trained on a parameterized 2D cylinder flow field dataset and using MeshGraphNet.
You can use the inference script to rollout the autoregressive model for new test cases and make an animation for comparing the predictions against the ground truth. Refer example for more details.
The model takes a dgl graph and nodes and edge features for that graph as the inputs. The node features include the velocity values at a specified time step, as well as a one-hot vector indicating the node type. The edge features consist of relative displacement vector and its norm.
Output is the predicted velocity and pressure at the next time step on the nodes of the input graph.