This is the result of interpolating experimental data on a finite element mesh using ResampleWithDataset filter.
The missing points along the circle have value zero (blue in the contour above) because these points were not measured or have NaN.
However, the distance to the points living on the circle is small, so I want to extrapolate (nearest neighbor, bilinear, …) to have values everywhere.
Here is a state file containing the data and the output of the filter state.pvsm (551.5 KB)
Here are the data (points and experimental values): nodes.csv (7.4 KB) data.txt (655.1 KB)
and the sequence of filters to generate the output
Is there any filter in paraview / pvpython that enables extrapolation or anything else I can do in paraview?
I could reproduce the result in your state file. Basically, you have added an selection u>=0 followed by another delaunay. The trick is that the selection u>=0 gets rid of the NaN’s in my data, hence forcing the delaunay to triangulate the entire domain resulting in interpolation everywhere, even between the hole. Is it like that?
ResampleWithDataSet provides a pickable option “Snap data to Cell with Closest Data”. Without choosing this option, points outside of the domain get a zero which for fields that cannot be zero is bad so I was very hopeful to see this option become available. However, to me, Snap means to not extrapolate, rather to take the closest boundary value directly.
Examining the data, it does seem to be extrapolating. According to the information tab which gives the Ranges → [ min, max]
Source
[0.0008571940043974022, 0.15147366645850846]
Target
[-0.0017042262373027119, 0.15147364187106566]
Note the source and target are generated from the same CAD so we are talking about roundoff error in the mesh generation that leaves a significant fraction of the points outside of the domain and critically lots of them on a wall surface with dramatic solution variation so this is really bad because the flow solver downstream will not accept 0 density as an starting field (equation of state blows up) and an extrapolated negative value is even worse.
I don’t see anyway for a target to give a negative value from a uniformly positive source field without doing extrapolation which to my understanding is not what snap means.
My hack around this, provides a “safe” but very inaccurate starting field by using transform to scale (shrink) and translate such that no points on the target mesh fall outside of the domain.
