Boundary Graph Neural Networks for 3D Simulations

Authors: Andreas Mayr, Sebastian Lehner, Arno Mayrhofer, Christoph Kloss, Sepp Hochreiter, Johannes Brandstetter

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The new BGNNs are tested on complex 3D granular flow processes of hoppers, rotating drums and mixers...BGNNs are evaluated in terms of computational efficiency as well as prediction accuracy of particle flows and mixing entropies. In this work, training, validation, and test data are generated by LIGGGHTS modeling particle trajectories within different machinery designs.
Researcher Affiliation Collaboration Andreas Mayr1, , Sebastian Lehner1, Arno Mayrhofer2, Christoph Kloss2, Sepp Hochreiter1,3, Johannes Brandstetter1, ,* 1 ELLIS Unit Linz & LIT AI Lab, Johannes Kepler University Linz, Linz, Austria 2 DCS Computing Gmb H, Linz, Austria 3 Institute of Advanced Research in Artificial Intelligence (IARAI), Vienna, Austria *now at Microsoft Research AI4Science
Pseudocode Yes Algorithm 1: BGNN: Dynamic Graph Message Passing
Open Source Code Yes Code is available at https://ml-jku.github.io/bgnn/
Open Datasets No In this work, training, validation, and test data are generated by LIGGGHTS modeling particle trajectories within different machinery designs. No specific link or citation for the generated dataset is provided.
Dataset Splits Yes In this work, training, validation, and test data are generated by LIGGGHTS modeling particle trajectories within different machinery designs.
Hardware Specification Yes LIGGGGHTS simulation is run on a CPU AMD EPYC 7H12, BGNN forward pass is run on a GPU NVIDIA A100.
Software Dependencies No No specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or other libraries) are mentioned in the text.
Experiment Setup Yes We use 3 to 10 message passing layers, with 128 and 512 nodes for intermediate node and edge representation. The cut-off radii strongly depend on the particle size. We use cut-off radii of 0.02 and 0.008 for rotating drum and hopper, respectively. Cut-off radii have been treated as hyperparameters of our model.