docs: S2 v0.5 - Gre24 methods mapping, V-dot target, B0 naive arm, R2-G24 features

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
ClaudeandClaude Opus 5.5 committed 2026-11-14 10:26:34 +00:00
1 parent 761d03c957
commit 1b55e16689
1 file changed
+44
@@ -0,0 +1,44 @@
# S2 design v0.5: Gre24 mapping and target revision (Claude, 2026-11-14)
**Source:** Greif, Lechner, Meyer, Composites Part A 185 (2024) 108340, author PDF from Augsburg OPUS docId 113920 (4.3 MB). Text was extracted locally with pypdf 7.1.0. Section numbers below are the paper's. Only short excerpts are paraphrased; the PDF is not redistributed.
## 1. What Gre24 actually does
- **Inputs (§3):** the tet mesh Ω, injection locations ∂Ω_i, and *the melt flow rate V̇ at the gate*. Per §2, V̇ is "determined automatically for each geometry by Autodesk Moldflow".
- **Distances (§3.1.1, Table 1):** eikonal ‖∇φ‖ = 1/f, solved with tetFIM (package fim-python) on the Moldflow mesh.
- D_w: f = 1, source = walls.
- D_i: f = 1, source = gate.
- D_f: f = D_w (wall distance used as a travel-speed proxy), source = gate.
- **Spatial moment (§3.1.2, eq. 6–7):**
- S(x_i) = Σ_j (v_j/V_i) d_ij ⊗ d_ij over cells j whose centre lies within R of node i and whose |D_i(cell) − D_i(node)| < R.
- R = 3·h_max (the longest element edge in the mesh).
- **Normalisation (§3.2, eq. 8–11):**
- D*_x = D_x / max D_x.
- Relative flow volume V*_f(c) = vol{D*_f < c} / vol(Ω), i.e. the volume fraction reached by the flow-distance front.
- Normalised gradients ∇*(·) = ∇(·)/‖∇(·)‖.
- **Feature tuple (eq. 12):** X* = (V*_f, ∇*V*_f, D*_i, ∇*D*_i, D_w, D*_w, ∇*D*_w, S).
- **Target (eq. 13, 17):** t*_fill, with denormalisation t_fill = (V/V̇)·t*_fill.
- **Models:**
- naive: t*_fill = V*_f;
- FNN: Adam, learning rate 1e-3, batch 1024 nodes;
- GBRT: XGBoost.
- **Evaluation (§4.1):** 5 random splits of 629 geometries, 315 test / 314 train, by geometry. Per-geometry RRMSE.
- **Fill-time result:** "ML models perform only slightly better than the naive approximation", and tuning does not notably beat naive. 300 of 314 RRMSE values are well below 0.1, and the median is about 0.04. The largest errors come from jumps in wall thickness.
## 2. Consequences for S2
1. **Target: revise the v0.1 B ruling.**
- Gre24 treats V̇ as a *process input*. Under Automatic control, V/V̇ is essentially the total filling time, so predicting seconds *without* V̇ makes the error dominated by one global scalar.
- That global error would swamp the local effect S2 is meant to test (the edge readings).
- **New primary:** t* = t·V̇/V, with V̇ the applied injection flow rate read from each part's LOG (a process setting, available at deployment), the same in all arms.
- Relative L2 is unchanged by a per-part constant scale, so seconds-level relative error equals t*-level relative error once V̇ is known. Only the oracle T_max normalisation stays forbidden.
- **To verify (MeltAstra):** the LOG contains the applied (velocity-phase) flow rate as a setting, separate from the time-series "Flow Rate" column. If it does not, fall back to the v0.1 seconds target and record why.
2. **Add arm B0 = naive t* = V*_f (Gre24 naive),** computed with our features. It is mandatory. Gre24 found the learned models barely beat it on fill time, so the S2 table must show E and N against B0 *and* R2.
3. **R2 becomes "R2-G24"** on Gre24's X* tuple, with documented deviations:
- D_w = our exact d(x) (the eikonal with f = 1 from the walls approximates it);
- D_i and D_f = tet-edge Dijkstra with edge cost = length (D_i) or length / mean D_w of the edge (D_f), not tetFIM, labelled "graph approximation of the eikonal";
- V*_f from D_f with lumped volumes;
- ∇* = normalised P1 gradients averaged to nodes;
- S with R = 3·h_max and the D_i filter, as in eq. 7;
- sklearn HistGBRT in place of XGBoost.
The v0.2 R2-G column list is superseded.
4. **Identifiability risk is now explicit.** If B0 already matches Gre24's reported accuracy on our grouped split, E vs N differences may be small. Then the six-region table reports "no meaningful gain" or "insufficient evidence", which is a legitimate result. The withdrawn 2 % stop is not reinstated.
5. **Split contrast:** Gre24 uses random geometry splits, while our S2 uses grouped components. Report a secondary random-split B0/R2 result only as a labelled "Gre24-protocol comparison"; it is not a primary result.