Voxelization routine

[Kresa96]

Very nice work!
We try to use your package MicroStructPy in our research group to built polycristalline electrolytes for impedance simulation. We usually use numpy arrays in which each grain is represented by several array entries (very similar to voxels, or images in 2D). The grains are distinguishable by different integer values in the matrix. In this way it's easy to modify the structure and use it for several different purposes (as an example for our simulations).

Example for a 2D structure consisting of 2 grains:
1111112222
1111112222
1111222222
1112222222

I think it would be very nice if you could implement a function which can transform the 2d and 3d polygon mesh into numpy arrays.

Looking forward to your reply！

[kip-hart]

Hi thank you! I'm glad to hear you're using MicroStructPy. I understand what you're asking for, where there is a grid of points and the value assigned to each point is the grain number, then these values are stored in a matrix.

I'm not familiar with voxelization approaches. My first thought is to create a list of points then test if point "i" is within grain "j", but that would be very slow. My second thought is to put bounding boxes around each grain, then only test if point "i" is within grain "j" if the point is within the bounding box. This is better, but probably still slow.

Do you have suggestions for voxelization routines?

[Kresa96]

Thanks for the fast anwer!
No I am also not very familiar with those methods. But I saw that a voxelization routine for ellipsoidal grains is already implemented in 'Kanapy', a Python package from 'mrgprasad', which was also made to create microstructures in Python. Maybe all you need is in their documentation.

In 2D I was able to translate the polymesh.plot in a very 'dirty' way into a numpy array:

polymesh.plot(facecolors = "white", edgecolor = "k")
plt.axis('off')
buf = io.BytesIO()
plt.savefig(buf, format="png", bbox_inches='tight' , pad_inches = 0, dpi=500)
buf.seek(0)
img_arr = np.frombuffer(buf.getvalue(), dtype=np.uint8)
buf.close()
img = cv2.imdecode(img_arr, 0)

s = [[1,1,1],[1,1,1],[1,1,1]]
labeled_mask, num_labels = ndimage.label(img, structure = s)
labeled_mask = expand_labels(labeled_mask, distance = 20)

But in 3D this dirty procedure will not work.
If you are able to implement something like this, we will probably use your package for lots of stuff.

[kip-hart]

Ok got it, I'll check out Kanapy. I've started a branch to add a RasterMesh class. The pseudo-code for creating the mesh is

MicroStructPy/src/microstructpy/meshing/rastermesh.py

Lines 113 to 122 in 555c61e

	# 1. Create node and element grids
	# 2. Compute element centers
	# 3. Remove elements outside domain
	# 4. For each region:
	# A. Create a bounding box
	# B. Isolate element centers with box
	# C. For each facet, remove centers on the wrong side
	# D. Assign remaining centers to region
	# 5. Combine regions of the same seed number (remove voids)
	# 6. Define remaining facets, inherit their attributes

This mesh class would have the same attributes as TriMesh. To create the array of grain numbers, my pseudo-code is

MicroStructPy/src/microstructpy/meshing/rastermesh.py

Lines 199 to 201 in 555c61e

	# 1. Create array full of -1 values
	# 2. Convert 1st node of each element into array indices
	# 3. For each element: populate array with element attributes

I'd also add functions to write these meshes out in VTK and Abaqus formats.

I can keep you posted with updates. How urgent is the need?

[Kresa96]

Nice!
Yeah I think it is a nice idea to implement the output of these files!
Keep me updated!
The need is very urgent 😁

[kip-hart]

I've added details in #45 for how this is implemented. You can get the desired array with a script that first creates/opens a PolyMesh, then calls RasterMesh.from_polymesh. The raster mesh created by this function has a method as_array. Array indices are 0 in the lower-left corner. The functions work in both 2D and 3D.

Upgrading to v1.5.0 will have these features. Please let me know if you encounter any issues.

[Kresa96]

Thank you very much for the fast implementation!!
I got a view problems:
In 2D the plot() method does not work in my case. If I plot the matrix with plt.imshow() for several values of the argument 'mesh_size' I get anomalies in the matrix:

In 3D voxelization does not work at all in my case:
Input:

Output:

[kip-hart]

Sorry to hear that. Could you attach the 2D and 3D polymesh.txt files? You can also email them to support@microstructpy.org if that doesn't work. I'll debug with those; it's likely a floating point error for the 2D case.

For the .plot() function, this is the default behavior when there are no options. If you assign a color to each seed, then .plot(color=seed_colors, index_by='attribute') should produce the desired result. You can also pass other matplotlib arguments, like .plot(..., edgecolor='k', lw=0.5). All the non-MicroStructPy keyword args are passed through to matplotlib. More details at https://docs.microstructpy.org/en/latest/api/meshing/trimesh.html#microstructpy.meshing.TriMesh.plot

[Kresa96]

Yes you are right the implemented 2D plot works for me:

But it seems like that the array gets turned and the anomalies arrise when as_array() gets applied to the grid.

I will send the files to the mail!
Thanks!

[kip-hart]

Since (0,0) is in the lower-left corner, plotting with plt.imshow(..., origin='lower') should prevent the turning. You can also use the extent=(left, right, bottom, top) option to scale the axes into physical space, rather than matrix coordinates

I'll debug the 2D plaid effect and 3D issues now and update you with results.

[kip-hart]

I've cleared up the plaid issue in 2D. It was a floating point error. Working on the 3D case now.