Deforming a mesh using MPM particles?

I’m using a triangular mesh to seed a shape in a MPM simulation. Is there a way of moving the mesh together with the particles? (I mostly need the mesh for rendering, but also for computing stuff like its curvature)
Thank you in advance!

I imagine @yuanming used something like that for generating the armadillo-man images? armodillo

1 Like

Hi! I’m not super sure how @yuanming made this cool image at that moment, but if you want to render 3D meshes, hope my extension library Taichi THREE, despite it’s still under-development, could helps :slight_smile:


thank you @archibate! will try it :slight_smile:

In my case, the starting point is a mesh. I used code from taichi_elements to voxelise the mesh and generate the MPM points. After that, my algorithm deforms the mesh, quite massively :sweat_smile:. At the end of the simulation, the points had move well far away from their original position. I would like to deform the mesh accordingly.
I was thinking to represent the position of each vertex in the original mesh as a linear combination of the k>=3 nearest MPM points. At the end of the simulation, I would compute the new position of the mesh vertices using the weights of the initial linear combination. Does this sound sensible to you? I’m afraid that computing k nearest neighbours with so many points and vertices may be super slow :snail:

Yes, that’s indeed super slow, but you may use a neighbour lookup table or octree for quicker lookup, e.g., pseudo-code:

for i in pos:  # scan neighbours
  x, y = int(pos[i])
  ti.append(neighbour, [x, y], i)

Hello! I found a way of deforming the mesh, and it works like a charm :smiley:. I thought I would share it here in case it can help someone else.
Here’s my setup: I have a mesh with vertices v and faces f, which I use to seed particles p. The particles get deformed, and turn into q. And what I want is to move the vertices v following the deformation of p into q.
There’s two steps: first, i compute k neighbouring points p for each mesh vertex v (using a kdtree). Second, I use moving least squares to compute a deformation using the k p neighbours of each vertex as control points.
Here’s the complete algorithm:

import numpy as np
from sklearn.neighbors import KDTree

# you need to get the following numpy arrays
# v: mesh vertices
# p: initial particle positions
# q: final particle positions
# The new vertices will be in V
tree = KDTree(p)
dist, ind = tree.query(v, k=k)
alpha = 1
w = np.array(1/dist**alpha)
V = v.copy()
for index in range(len(v)):
    pcentroid =
    qcentroid =
    ptr = pi - pcentroid
    qtr = qi - qcentroid
    M =
    detM = np.linalg.det(M)
    if detM < 1e-8:
        V[index] = vi - pcentroid + qcentroid
        invM = np.linalg.inv(M)
        for j in range(5): # k = 5
            res = (vi-pcentroid).dot([j]*ptr[j]))
            A[index,j] = res
        V[index] = np.sum(np.diag(A[index,:]).dot(qi),axis=0) + qcentroid

Finally, here’s pictures of one of my brain meshes before and after deformations.