Recent advances in message-passing PDE solvers
Please login to view abstract download link
The interest in learned simulators, given their fast inference time, has grown in the last times. Amont them, those based upon message-passing Graph Neural Networks have become very popular, due to their formal similitude to any mesh-based technique (finite elements, finite volumes, finite differences) and their ability to generalize to previously unseen meshes. However, they have one fundamental drawback: they scale very poorly. Their memory consumption is, in general, high, and state-of-the-art meshes often become difficult to parse. In this work we analyze alternatives to the classical MeshGraphNet architectures and provide with promising solutions to the afore mentioned limitations.