GPU Module Unexpected Behavior

[David Reger]

Hello,

I have been running some granular flow simulations with the GPU module. When comparing my results to other similar work that uses other codes, my results are not aligning with those from the other works, even when the same material properties are used. I cannot post my results publicly on the forum, but can share them with someone from the Chrono team over email. 

Essentially, I am seeing that in previous works, the streamwise velocity of particles is continuously decreasing from the center of the granular bed towards the wall. In my simulation, however, the velocity of particles is essentially the same across the bed, except for the particles that directly touch the wall.

The paper I am comparing to is:

X. Sun, Three-Dimensional numerical simulation of quasi-static pebble flow

In particular from this paper in Figure 9, my results look much more like the Hookean velocity profiles than the Hertzian profiles that agree with experimental results. Any help on understanding what is going on would be appreciated.

Thanks!

David

[Ruochun Zhang]

Hi David,

This is a very involved question because, I imagine that only people with research experience of pebble flow problems can confidently step forward and claim to offer "technical insights". I have never researched said problems. There are a number of suggestions however, that Luning and I think we can offer.

First please make sure the scale/resolution of the simulations matches. The number and grain size of particles can make a big difference. It seems the data in this paper is measured under certain equilibrium conditions, that is also something to pay attention to.

Then perhaps more importantly, I am pretty sure these 2 implementations do not use the exact same contact model. The normal damping term seems to be slightly different and, it seems this paper did not mention how the tangential stiffness and damping coefficients should be calculated. This could be key information: in this quasi-static problem, the tangential force is directly in the "downward" flow direction and should play an essential role. This is true even when you have the same friction coefficient 0.3, because 1) the normal force may be different; 2) I imagine for many contact pairs this tangential force is in the "micro displacement" region and not equal to 0.3 times normal force. Also, I didn't quite follow the explanation for the differences between the Hookean and Hertzian models in the paper, that supposedly caused the discrepancy we see in Fig. 9. How are they different in the paper? In short, it's quite likely we are not using the same force models, even though they are both Hertzian model; but just from this paper we cannot verify that.

Following the previous argument, from a quick read-through I don't know if it's indeed because of the discrepancy between Hookean and Hertzian, as opposed to the discrepancy between  the exact models Rycroft and Sun et al used. Again, maybe it is well-known in that research community, it's just that I don't know. My engineering intuition though, is that this should be similar to a CFD boundary layer problem, where you see sharp near-boundary velocity decline for non-viscos flows, and "flatter" velocity profiles for viscos flows. And following this line of thinking, you may be able to match the velocity profile that you are after, by correctly modelling the flow "viscosity". Maybe you can try adding cohesion (cohesion coefficient of something around 10) and/or some rolling resistance and/or modified tangential stiffness, to see if it achieves something similar. You probably would not use this "empirical solution" in the end, but it can be a sanity check, so we know whether some "subtle" settings can be useful in capturing this physics, even as useful as using a particular class of model.

I noticed many DEM users call for "slightly customized" force models, that is something I try to reflect on the APIs of the upcoming next-gen Chrono DEM solver. But with current Chrono::GPU, you have to go to function materialPropertyCombine to modify the material-property-to-normal-tangential-stiffness conversion, and to go to function computeSphereContactForces_matBased to modify contact force calculations, then rebuild the project. You could do that and maybe we can help you do that, if a clear definition of the force model is given, and you would like to reproduce that result.

I cannot make comments specific to the pebble flow problems, and someone with the research experience can definitely do a better job; but I hope this helps. 

Thank you,

Ruochun

[David Reger]

Hi Ruochun,

Thanks for your insight! I suspect that the difference in contact models is likely the discrepancy between my work and the other paper. I'll see if I can implement the contact model that they had used and if it will improve my results. I'm currently working to make a model that aligns better with another experiment to perform a more proper validation exercise.

Thanks,

David