MeshMask: Physics Simulations with Masked Graph Neural Networks

ICLR 2025 Submission

Code and Datasets: to be released

The proposed Masked Graph Neural Network architecture. During pre-training, we remove a portion of the edges and nodes of the mesh (e.g. 50%) and use this as input for the Encoder. Masked nodes are then replaced by a [MASKED] token before the mesh is passed to the Decoder. After pretraining, only the Encoder is finetuned on a specific dataset, without any nodes being removed.

cylinder-masked.mov

Cylinder

(top): Masked Input

(middle): Reconstructed Input

(bottom): Original Input

bezier-masked.mov

Bezier

(top): Masked Input

(middle): Reconstructed Input

(bottom): Original Input

3D-aneurysm-masked.mov

3D Aneurysm

(top): Masked Input

(middle): Reconstructed Input

(bottom): Original Input

compar-205.mov

2D Aneurysm

(left): Ground Truth

(right): GNN Prediction

compar-160.mov

2D Slices from 3D Predictions

(left): Ground Truth

(right): GNN Prediction

compar-128.mov

compar-tawss.mov

Time Average Wall Shear Stress

(left): Ground Truth

(right): GNN Prediction

3D Aneurysm Comparison

Comparison between CFD and our GNN on the plane y=10 and 3 selected points inside the aneurysm.

Page updated

Google Sites

Report abuse