Soft body dynamics for medical interactive simulation

[davide445]

Was considering MFEM for a medical interactive simulation project.
Searching a bit I didn't find any usage related to this sector, so wanted to ask how suitable are current solvers for soft body (human tissue) dynamics realtime simulation.
This will need to be integrated with biofluids simulation and perfusion, haptic feedback and immersive visualization (will be used a library with all compute based rendering pipeline to minimize latency).
There are more specialized frameworks such as SOFA, but we are targeting medical grade accurancy and AMD acelerators that are more in line with MFEM strengths.

[v-dobrev]

Hello Davide,

As far as I know, soft body tissue models are similar to nearly incompressible non-linear elasticity. If that is indeed the case, we have two examples that might be of interest for you to look at: ex10/ex10p (time-dependent Neo-Hookean non-linear elasticity) and ex19/ex19p (steady-state incompressible nonlinear elasticity) -- you can see short summaries of these and other examples/miniapps at https://mfem.org/examples.

Porting and optimizing these to GPUs will require some work, however, the elasticity kernels are of interest to other projects and I think we'll have initial support for elasticity on GPUs very soon. Also, such kernels are very similar to the TMOP (mesh optimization) kernels we have already ported to GPUs.

Regarding the integration with the other components you mentioned (like biofluids simulation and perfusion), do you envision using MFEM for those too or you have other tool(s) in mind? This sounds like a coupled fluid-structure interaction simulation which is also of interest to us, even though we currently do not have an example or miniapp like this.

Other project members can probably add to my summary: @tzanio, @jandrej, @jamiebramwell.

-Veselin

[davide445]

Hello
we are prototyping with various methodologies, included ML.
As you wrote approximation is always useful to speedup things, the point for me is expand the computational capability of your architecture instead of compromising the quality.

[BBBBBbruce]

Yes, I agree, when the thing comes to the Biomedical area, quality is always the top priority. Have you tried PhysX 5.0? They have GPU supported Realtime FEM implementation. It might be useful.

[termi-official]

Since I am working on speeding up mechanics simulations, let me share my experience here. Projective dynamics has a rather limited capability - Tiantian Liu has developed a clever Quasi-Newton method (linear elasticity surrogate), but it also comes with the issue of how to costruct a suitable Jacobian, especially for advanced material models. For real time your best bet may be physics-informed neural networks (or in general techniques from the PDE-constrained zoo of technique) to derive a surrogate, unless you have really small number of elements and simple material models. You can use some offline FEM simulations to generate your training data.

[davide445]

For real time your best bet may be physics-informed neural networks

Of course this is a path we are exploring

You can use some offline FEM simulations to generate your training data.

The problem here is (apart the key scientific and medical aspects to setup a realistic enough simulation) as always the computational effort and elapsed time to generate a detailed training set for an high resolution simulation. This is one of the reasons I'm interested into MFEM.

[termi-official]

Great, I wish you good luck with this appraoch!

Generating the data is ideed an issue. However, PINN's seem to have a very interesting property here. In comparison with learning decision making or vision problems, experiments which suggest that we do not need a super large data set to obtain a good set of weights (because we impose much structure through the PDE constraint). Note that the Networks also tend to be smaller, which might play a role. But I want to note that I personally do not have experience with PINN's beyond fitting surrogates for toy problems.