search

brandstetter-johannes / MP-Neural-PDE-Solvers

Repo to the paper "Message Passing Neural PDE Solvers"
MIT License
121 stars 24 forks source link

Artifacts with Equation 3 #14

Open LouisSerrano opened 1 month ago

LouisSerrano commented 1 month ago

Hi, it's me again. I am trying to generate data for the combined equation (E3) and I am struggling to obtain trajectories without aliasing artifacts / divergence for parameters approaching the KdV equation (alpha close to 3, gamma close to 1 and beta close to 0.1). I am using the default hyper parameter range and a grid resolution of size 256.

brandstetter-johannes commented 4 weeks ago

Hey Louis, what exactly do you mean by aliasing effects. Very likely that the data generation is sub-ideal. Can you post an example?

LouisSerrano commented 4 weeks ago

Thanks for the reply. Sure, I have this screenshot available, it is not the most unstable generation, but it can give you an idea of the issue. It is hard to predict when the solver will produce this kind of artifacts and it depends heavily on the initial condition, here it is with parameters $\beta=0.19, \alpha=1.15, \gamma=0.87$.

Instabilities_CE
LouisSerrano commented 4 weeks ago

I can post more examples if you want, or in another format if it can help.