Makie surface and NaN values

[vettert]

I am running in some unexpected behavior with GLMakie, surface plots and matrices that contain NaN as a value.

My minimum working example below produces the following plots of three different 4 x 4 matrices with different elements set to NaN:

1. What causes the darker shading of cells adjacent to a NaN value?
2. The right most example is, frankly, somewhat weird to me (why is the surface dropping to a low value and becomes transparent like this?)

The behavior that I would like to have is that all tiles that have a NaN value as one of its vertices are simply not displayed. Is this doable with some tinkering of attributes of the surface?

Here the MWE code reproducing the above figures:

using GLMakie

x = 0.0:3.0
y = 0.0:3.0
Z1 = reshape([100*i+50*j for i in x for j in y], (length(x), length(y)))
Z2 = copy(Z1)
Z3 = copy(Z1)

Z2[1,1] = NaN
Z3[1:2, 1:2] .= NaN

f = Figure(resolution = (800, 800))
ax1 = Axis3(f[1, 1], aspect = (1, 1, 1))
ax2 = Axis3(f[1, 2], aspect = (1, 1, 1))
ax3 = Axis3(f[1, 3], aspect = (1, 1, 1))

surface!(ax1,x,y,Z1)
surface!(ax2,x,y,Z2)
surface!(ax3,x,y,Z3)

[rafael.guerra]

Just for reference, linking a related issue here, for which did not get an answer.

[vettert]

The behavior in your post is actually what I want. Smooth edges would be nice, I agree with you on that, but I would actually settle for what you already had!

The problem is, however, that your code produces a completely different looking image on the most current Makie version. It now shows similar issues as are shown above, but since your z-values are negative, the surface doesn’t go down like in my right hand side image, but up (towards 0).

edit: Included an image for reference.

[rafael.guerra]

Empirical tip for the latest Makie version: replace NaN by Inf

PS: but result is still not shaved though…

[ffreyer]

The shader computes normals using the surrounding grid positions. If those have weird values the normals will have weird values and the lighting calculation will do weird things. Setting shading = false should get rid of the dark rectangles (or alternatively setting diffuse=Vec3f0(0), specular = Vec3f0(0)).

[rafael.guerra]

@ffreyer, fyi none of your solutions worked for me. The result is still as displayed by @vettert for the truncated paraboloid example.

[ffreyer]

using GLMakie

x = 0.0:3.0
y = 0.0:3.0
Z1 = reshape([100*i+50*j for i in x for j in y], (length(x), length(y)))
Z2 = copy(Z1)
Z3 = copy(Z1)

Z2[1,1] = NaN
Z3[1:2, 1:2] .= NaN

f = Figure(resolution = (800, 800))
ax1 = Axis3(f[1, 1], aspect = (1, 1, 1))
ax2 = Axis3(f[1, 2], aspect = (1, 1, 1))
ax3 = Axis3(f[1, 3], aspect = (1, 1, 1))

surface!(ax1,x,y,Z1,shading=false)
surface!(ax2,x,y,Z2,shading=false)
surface!(ax3,x,y,Z3,shading=false)

This does get rid of the darker rectangles for me. Changing shading requires you to close the window beforehand though.

[ffreyer]

The reason for the surface to drop like it does is that the shader converts nan values to 0. You could work around that by passing the modified z-values as colors, i.e.

surface!(ax3,x,y,Z1,shading=false, color = Z3)

I guess for a smooth cut we’d need a shader discarding pixels that are outside some value range or more generally a shape

[vettert]

Thanks to both of you @ffreyer and @rafael.guerra. Here is the full code that gets close to what I wanted.

using GLMakie

x = 0.0:3.0
y = 0.0:3.0
Z1 = reshape([100*i+50*j for i in x for j in y], (length(x), length(y)))
Z2 = copy(Z1)
Z3 = copy(Z1)
Z4 = copy(Z1)
Z5 = copy(Z1)


Z2[1,1] = NaN
Z3[1:2, 1:2] .= NaN
Z4[1,1] = Inf
Z5[1:2, 1:2] .= Inf

f = Figure(resolution = (800, 800))
ax1 = Axis3(f[1, 1], aspect = (1, 1, 1))
ax2 = Axis3(f[1, 2], aspect = (1, 1, 1))
ax3 = Axis3(f[1, 3], aspect = (1, 1, 1))
ax4 = Axis3(f[2, 2], aspect = (1, 1, 1))
ax5 = Axis3(f[2, 3], aspect = (1, 1, 1))

s1 = surface!(ax1,x,y,Z1)
s2 = surface!(ax2,x,y,Z2)
s3 = surface!(ax3,x,y,Z3)
s4 = surface!(ax4,x,y,Z4, shading = false)
s5 = surface!(ax5,x,y,Z5, shading = false)

display(f)

The smooth cut thing is of course nice to have, but in my use example, I am fine with just making the grid finer and minimizing how well one sees the rugged edges (the visualization is not the computationally expensive part of the process )

[rafael.guerra]

NB: using Inf workaround as indicated further above.