search

tum-pbs / SFBC

ICLR'24: Symmetric basis convolutions for learning lagrangian fluid mechanics
MIT License
5 stars 2 forks source link

Running on GNS data #1

Open iSach opened 1 month ago

iSach commented 1 month ago

Hello,

First of all, thank you for releasing the code for your great paper. I am currently looking into benchmarking SFBC on the datasets from the paper "Learning to simulate complex physics with graph networks" aka GNS. For example, I am mainly trying to make the model learn to model the physics of "WaterRamps", which consists of blocks of water dropped in a square domain with rigid boundaries, and solid ramps within the domain.

The only data available is pretty much the positions and velocities. I benchmarked DMCF on the GNS data as it contains utils for that, and fits the problem formulation pretty well and does not require much hyperparameter tuning. However, I am struggling to make SFBC work on this dataset.

I tried to convert the dataset into the format of the case II dataset, by filling missing hyperparameters/inputs. However, even though the model is now training, it does not learn anything.

image

Have you tried running SFBC on such data? I would like to know how it compares to GNS and DMCF for emulating such datasets. Do you maybe some indications and advice for using it on these datasets?

Thanks in advance, Sacha

iSach commented 1 month ago

Here is the script used for conversion: https://pastebin.com/aERJRjRX

Please note that we took the parameters from test case II.