you are touching an intersting and problematic topic of triangualted surfaces. In DEM, usually particles have only 2-contacts, and don't know about topology of other particles; plus, particles deformations are modeled as small overlaps. Now imagine a sphere which is lightly overlapping with triangulated surfaces rolling over the edge between two facets: it will suddenly be pushed upwards by 2 contacts (instead of one) and the edge will push it also sideways (very slightly).
I wrote a script to demonstrate that: the trace is scaled vertically 10000x (the overlaps are really very small) and you see that there are 2 contacts around the common edges, and that's where the center moves upwards, then falling again and so on. Let's see if embedded videos works:
OK, they do, but youtube only plays this with 360px quality, horrible; this is an image of that:
Now, is there a way around that? You bet there is, but it is not implemented (though I though about it a lot, and some products, like PFC, have an optional feature called "smooth surfaces" which AFAICT does just that). It incurs some performance penalty, as contacts are not independent on each other (in each step, you have to solve all contacts of one particle first, then somehow decide how to put them together and so on; that implies worse parallelism results), and corner cases are really hard to do: common edge is easy, but now what you do when a sphere rolls on the edge, and comes to a vertex common to e.g. 6 facets which are not co-planar? Smoothness of the solution (no abrupt change of normal, or of force between consecutive steps) is a requirement here.
What you see is a result of that. The usual practice is just to think, ok, we know it is not perfect, but it has only a small effect on the global result. If you need more (e.g. discretization-independent contact with the surface), you have hard things to solve ahead of you.