Questionable results of stokes mgn from nvidia modulus

[malaika]

Hi, i reproduced some results of stokes mgn example in nvidia modulus, visualized them in paraview, at different scales they look good but within the same range questionable, also the error curve was flat at the end of training also inference so its not related to time,
at different scales,

[malaika]

continued…

at same scale,

so was it supposed to be this way?

[malaika]

filtered are after physics informed fine tuning,
pred are from gnn inference