MeshMask: Physics-Based Simulations with Masked Graph Neural Networks

Paul Garnier, Vincent Lannelongue, Jonathan Viquerat, Elie Hachem · 2025

Read MeshMask: Physics-Based Simulations with Masked Graph Neural Networks by Paul Garnier, Vincent Lannelongue, Jonathan Viquerat, Elie Hachem (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-08738.txt.

Description

Paul Garnier, Vincent Lannelongue, Jonathan Viquerat, and Elie Hachem pretrain graph neural networks for fluid simulation by hiding parts of the mesh. Masking up to 40% of input mesh nodes forces the model to learn sturdy representations of fluid behavior. An asymmetric encoder-decoder design and gated multilayer perceptrons add power. Tests ran on seven computational fluid dynamics datasets.

One new dataset simulates 3D intracranial aneurysms with over 250,000 nodes per mesh. The method set leading results across the suite and lifted long-term prediction accuracy by up to 60% over prior best models at like cost. Pretraining on several datasets at once also cut the time and data needed for new tasks.

Catalog record

CREATORPaul Garnier, Vincent Lannelongue, Jonathan Viquerat, Elie Hachem
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-08738

Browse the full library · About the Tower of Babel