This is not an official blogpost. It is only supportive material for our ICLR 2022 submission Boundary Graph Neural Networks for 3D Simulations.
BGNN Example Trajectories
Overview
- Models with a Cohesive Material
- Models with a Non-Cohesive Material
- Failures
- Out-of Distribution (OOD) Experiments
Models with a Cohesive Material
Drum
Successful trial of training a BGNN for a cohesive material within a rotating drum (250.000 simulation time steps):
Hopper
Successful trial of training a BGNN for a cohesive material within a hopper device (100.000 simulation time steps):
Models with a Non-Cohesive Material
Drum
Successful trial of training a BGNN for a non-cohesive material within a rotating drum (250.000 simulation time steps):
Hopper
Successful trial of training a BGNN for a non-cohesive material within a hopper device (100.000 simulation times teps):
Failures
Drum
Failure trial of training a BGNN for a cohesive material within a rotating drum (250.000 simulation time steps):
Hopper
Failure trial of training a BGNN for a cohesive material within a hopper device (100.000 simulation time steps):
Comment on Failures
Local (small) modeling errors of particle-wall interactions result in a leak in the wall and some particles incorrectly cross the wall, whereafter many other particles follow through the leak
Out-of Distribution (OOD) Experiments
Drum
The upper animation below shows a typical training example, while the animation directly beneath is an OOD example.
For OOD test samples the length of the cylinder has increased compared to training and test samples.
Hopper
The upper animation below shows a typical training example, while the animation directly beneath is an OOD example.
For OOD test samples the outlet hole size has decreased and the inclination angle has increased.