What we found
Notes are short and dated — something we hit during real work, written down while the numbers were still on the machine. Papers are preprints with a permanent DOI and a published dataset you can check.
These are simulator results. The solver notes measure how faithfully a numerical method reproduces itself — finer grid against coarser, learned operator against the solver that taught it. No result on this page has been validated against a molded part on a press. Where that matters to a finding, the note says so in its own opening lines.
Findings
Pitch convergence of a freeze-margin field depends on fill time, not wall thickness
Refining a voxel fill solver 2× keeps the freeze-margin ordering on parts that fill in 0.15 s or longer (Spearman 0.93) and destroys it under 0.04 s (0.39) — and wherever the direction can be checked, refinement moves the field warmer, never colder.
More geometry beats cleaner labels
Retraining only on the half of the corpus whose labels are pitch-converged makes the operator worse on that very regime (0.939 vs 0.955), and down-weighting the rest changes nothing — at two thousand parts it is data-limited, not label-limited.
A learned pitch correction loses to upsampling; a tile head needs only the entry face
A residual operator trained to turn a coarse fill into its 2× version scores below plain upsampling (0.767 vs 0.830); and a tile head matches the whole-part operator using no melt-state information at all.
Where a coarse fill solution is wrong: its own cold cells, not its thin walls
Local wall thickness predicts almost nothing about where a coarse fill disagrees with its refinement (Spearman 0.10), while the coarse solution’s own value predicts it at 0.87 — so a refinement criterion has to be computed after a coarse pass.
Through-thickness resolution and injection pressure in a 3-D voxel fill solver
Predicted peak injection pressure changes by a factor of 4.0 between 3 and 6 cells across a 2 mm wall, with no sign of convergence — and a pressure cap silently censors the very quantity being converged.
The control-volume filling timestep has an unstated precondition
The standard filling timestep rule assumes a front scheme that snaps filled cells to exactly 1.0. On non-snapping advection it degenerates — plus a global-vmax CFL pathology and a diagnostic that overstated the opportunity by 267–808×.
Geometric multigrid fails on one-cell-thick shells, and two bugs it takes to find out
A V-cycle worth 9× on a solid box is a regression at every resolution reachable on real part geometry, because a one-cell wall cannot be coarsened.
Schur-complement degradation on molded-shell geometry
Outer iteration count grows 116 → 214 on a shelled tray where it is flat on a solid cube. The remedy the literature prescribes helps, and helps less as resolution rises.
How we publish research
What gets published here, how customer material is cleared, and where AI assistance is used.