CHAPTER 30
Advanced Solver Parameters
Engineering controls for experienced users - separate from the basic flow so as not to overload the beginner.
The parameters below affect performance and numerical stability. The defaults are starting points for most cases; adjust them only with a clear objective and validate the effect on the quantity of interest.
30.1 Neighbor Search Parameters
| PARAMETER | EFFECT | STANDARD / RECOMMENDATION |
|---|---|---|
| Verlet-list margin (skin) | Neighbor-list margin; balances rebuild frequency against candidate count | ~1⁄2 of the maximum radius (Chapter 12.4) |
| Rebuild frequency (GPU) | Number of steps between GPU list rebuilds | A larger interval reduces rebuilds but requires checking neighbor-list validity |
Contact/neighbor structures grow automatically; fixed slot tiers are not a user setting. Treat allocation failures as memory limits: inspect the log, particle/contact count and neighbor margin before reducing physical resolution. Model configuration is in Chapter 10; quantitative verification cases are in Appendix H.
30.2 GPU precision and execution mode
- Mixed precision (FP32) - computes contacts in FP32 and enables Verlet automatically. Turn it off for calibration and CPU comparison;
30.3 Advanced numerical diagnostics
Symptoms and responses to problems that appear in large or aggressive cases:
| SYMPTOMS | PROBABLE CAUSE | ACTION |
|---|---|---|
| Non-finite values (NaN) / overflow | Δt too large, initial overlap, incompatible stiffness | Reduce Δt; check particles born within geometry; review E and kn |
| Kinetic energy grows without cause | Numerical instability | Halve Δt; review damping and overlaps |
| CPU and GPU diverge in trajectory | Summation order and physical chaos (expected) | Compare trends/statistics; fix backend and threads for reproducibility |
| Slow case for no apparent reason | Inefficient Neighbor Search / High Save Count | Activate Verlet; adjust margin; reduce output frequency |
Change one parameter at a time - When adjusting advanced parameters, change one at a time and rerun a short interval comparing the kinetic energy curve and the quantity of interest. Simultaneous changes make it impossible to attribute cause and effect.
