Difftaichi for inverse simulation

Hi folks,

I am a new entry into the field of simulation and differentiable physics simulators. My experimental setup involves the following:

  1. A simulation loop that does the forward pass of a randomly initialised model of a deformable material using FEM.
  2. The deformation from the FEM sim is used to estimate an error.

The idea is to converge to a set of material parameters that reflect the actual real world material.

I was wondering if difftaichi could be used to invert FEm simulation for deformable materials? Could you point me to some resources/papers where people may have done so.


Maybe this is what you want:


In the first link, in the file of fem-explicit.py, the code uses the difftaichi. And the second link is some slides.

Hi, this is an interesting question, and you can check difftaichi.

I am wondering when numerical optimization iterations are involved in FEM simulation (e.g. Solving KU = F or minimizing the energy of the system) and how would you make this optimization process differentiable?

1 Like