CHAPTER 28

Model Limitations and Validity

What BrixDEM assumes, where the model is approximate and how to circumvent each limit.

Knowing the boundaries of the model avoids wrong conclusions. The limitations below are intrinsic to solver modeling choices - none prevent useful engineering analyses, provided they are interpreted correctly.

28.1 Particle shape

Particles may be spheres or rigid bodies composed of spheres (clumps, 10.8). For spheres, rolling and torsion can calibrate part of the real-shape response without reproducing geometry. Clumps explicitly represent elongation and angularity, with pebble-limited resolution and no internal rigid-body deformation. Choose and calibrate representation for the phenomenon of interest; screening, interlocking or orientation may require clumps and dedicated tests.

28.2 Average quantities measured on surfaces

The measured quantities on STL meshes (force, pressure, wear) are accumulated by triangle. The reported mean is a simple average over triangles. This is not a defect - it is a calculation convention: in meshes with approximately uniform triangle sizes, it represents the surface well. Only in meshes with triangles with very different areas can the simple average be less representative than an area-weighted average; in this case, also analyze the maximum value and spatial map (colorization by triangle), which do not depend on this weighting.

28.3 Total energy reported

The "total energy" magnitude of the standard output adds kinetic energy and gravitational potential. It does not include all elastic and adhesive energy stored in contacts (normal/tangential springs, JKR bond). In cases with cohesion or very rigid contacts, use energy as a trend indicator (stabilization, divergence), not as a closed energy balance.

28.4 Determinism and CPU × GPU differences

CPU and GPU implement the same physics, but the order of force summation differs between backends and thread counts. Small numerical differences are expected and are amplified in chaotic systems (many collisions). Compare aggregate trends and statisticsNot particle-to-particle trajectory. For exact reproducibility, fix the backend and number of threads.

Beyond the rounding difference above, there are three practical points when switching backend:

AspectCPUGPU
Physics modelsHertz–Mindlin, LSD, rolling/torsion, JKRThe same (model parity)
Numerical precisionDouble (FP64)Double; optional mixed precision computes contacts in FP32 and enables Verlet (faster, may cause energy drift in long or sensitive cases)
DeterminismDeterministic with a fixed thread countDeterministic with backend and settings fixed
JKR wall cohesionFull JKR at the wall too: bonded state, tension across an open gap and rupture at pull-off, with R_eff = R (double the P–P force). Verified in Appendix H, Case 8b

The comparison above concerns the cited models and cases; it does not establish parity for every combination, especially liquid bridges and breakage. Check case-specific coverage in Appendix H. For performance, the GPU only pays off with enough parallel work (see 27.2).

28.5 Low restitution requires model calibration

The viscoelastic damping used by the solver (Chapter 3) does not make the effective restitution match the prescribed value exactly across the whole range — the analytical verification in Appendix H (Section 2) shows that, at low e, the damping produces an effective e slightly above the prescribed one, and this relative deviation grows the lower the prescribed e is. For restitution coefficients prescribed below ≈ 0.5 (fairly inelastic materials — wet powders, soft pellets, ore with high energy absorption), do not trust the nominal value entered on the material/interaction: calibrate and observe the effective e via the particle drop test (Chapter 18) before using the result quantitatively. This applies to both Hertz–Mindlin and Linear Spring–Dashpot — for the latter, also remember that damping is not derived from e automatically (Appendix H, 11.2): enter a damping value consistent with the desired e.

28.6 Mixed precision and parity

Hertz, LSD, primitives, clumps, STL and JKR tests show close agreement across mixed-precision GPU, double-precision GPU and CPU. Threshold-dependent quantities can show a small systematic bias: in the measured cases, about 1–2% for angle of repose and flow. These are results for the tested cases, not universal tolerances. For friction or rupture calibration, compare the quantities of interest with mixed precision off.

28.7 Liquid bridge and breakage compatibility

Model availability, equation verification and material validation are different conclusions. Configure models in 10.9–10.10 and consult Appendix H cases/results for the backend and conditions actually checked.

CombinationUse and limit
Liquid bridges with spheres; particle–particle or particle–wallModel for small, isolated bridges. High content, coalescence and saturation are not checked by the bridge volume limit. Check force and inventory separately.
Liquid bridges with clumpsUnsupported; the GUI rejects liquid content in a clump material and liquid bridges in a pair with clumps. This does not prevent JKR/linear cohesion in compatible pairs.
Ab-t10 with PRM, without liquid bridgesCan replace a spherical or clump parent with spherical fragments. Clump orientation and contacts require calibration; an equivalent-diameter preview does not reproduce a clump DWT.
Prescribed bridges with breakage/fragmentationSupported. Breakage preserves prescribed bridges (it does not convert them into film), checked on CPU, GPU FP64 and mixed precision. This verifies the liquid balance; the behavior of a wet material with breakage is not experimentally validated.
Liquid migration with breakagePassing film to children does not establish validity of the complete interaction. Isolated-case coverage does not automatically validate this combination quantitatively.
Coarse-grainingLiquid parameter scaling is described in 10.7; its verification does not establish macroscopic equivalence of every process. Ab-T10/PRM breakage requires CGF = 1: the GUI and the solver reject any other value.
CPU, GPU FP64 and mixed precisionParity conclusions apply only to tested models and cases. A CPU execution of a CUDA-built DLL does not verify GPU kernels. See each case’s coverage in Appendix H.