From 3193734a7102c5bb648b2447f6eae2c3db7467c1 Mon Sep 17 00:00:00 2001 From: Claude Date: Sun, 15 Nov 2026 15:05:09 +0000 Subject: [PATCH] Gre24 faithful reproduction spec and tonight's plan (from author PDF) Co-Authored-By: Claude Opus 5.5 --- .../gre24-reproduction-spec-2026-11-15.md | 46 +++++++++++++++++++ 1 file changed, 46 insertions(+) create mode 100644 docs/claude/gre24-reproduction-spec-2026-11-15.md diff --git a/docs/claude/gre24-reproduction-spec-2026-11-15.md b/docs/claude/gre24-reproduction-spec-2026-11-15.md new file mode 100644 index 000000000..8f6e4b709 --- /dev/null +++ b/docs/claude/gre24-reproduction-spec-2026-11-15.md @@ -0,0 +1,46 @@ +# Gre24 faithful reproduction: specification and plan (Claude, 2026-11-15 night) + +Source: Greif, Lechner, Meyer, Composites Part A 185 (2024) 108340, author PDF from OPUS docId 113920, full text extracted locally with pypdf (scratch_pdf/gre24.txt). Section, equation and table numbers refer to the paper. R2-G24 in S2 was an approximation and is not this reproduction. The frozen S2/S3 results stay unchanged. + +## 1. Method as published (status: S = specified, A = ambiguous, M = missing) +| Item | Paper | Status | +|---|---|---| +| Data | Autodesk Moldflow simulations of 629 random ABC geometries; one random gate; PP GF30; automatic flow. Dataset DOI in footnote 1: **10.5281/zenodo.10027027** (v4 era); "Data will be made available on request" | S. Check whether the v4 record is downloadable | +| Mesh | Reuses the Moldflow tet mesh; all nodes | S | +| Distances (Table 1, eq. 5) | Eikonal on the tet mesh with **fim-python (tetFIM, Fu et al.)**, phi0 = 0. D_w: f = 1, source = wall boundary. D_i: f = 1, source = gate. D_f: **f = D_w**, source = gate | S. Gate source set (node or face) is A | +| Normalisation (eq. 8) | D*_w = D_w/max D_w, D*_i = D_i/max D_i, D*_f = D_f/max D_f | S | +| Relative flow volume (eq. 10) | V*_f(c) = vol{D_f < c}/vol(Omega), evaluated at each node's own D_f | Discretisation (nodal lumped vs element volume, ties) is A | +| Gradients (eq. 11) | Normalised gradient direction grad/norm of V*_f, D*_i, D*_w | How nodal gradients are formed on tets is A | +| Spatial moment (eqs. 6-7) | S_i = sum_j (v_j/V_i) d_ij outer d_ij over cell centres j with distance to x_i < R **and** abs(D_i(x_j) - D_i(x_i)) < R; R = 3 h_max (longest edge in the mesh); scikit-learn BallTree | S. D_i at cell centres is A; component layout is A (6 unique) | +| Feature tuple (eq. 12) | X* = (V*_f, grad*V*_f, D*_i, grad*D*_i, D_w, D*_w, grad*D*_w, S); D_w enters un-normalised as well | S (about 19 columns) | +| Target / denormalisation (eq. 17) | t*_fill = t_fill V-dot / V; t_fill = (V/V-dot) t* with the applied flow rate V-dot | A for us: does v4/v5 carry V-dot, and does V/V-dot equal our t_set? | +| Naive model (3.3.1) | t*_fill = V*_f | S | +| GBRT (3.3.3, 4) | **XGBoost**, 100 trees, max depth 8, learning rate 0.1, MSE | S. Other XGBoost parameters are library defaults, A | +| FNN (3.3.2, 4) | 4 hidden layers x 100, ReLU, Adam lr 1e-3, batch 1024 node pairs, MSE, PyTorch | Epochs / early stopping are M | +| Multi-output | Y* = (t*_fill, t*_cool, S*_vol, A); the paper trains on all of them | We reproduce fill time only. Single-output versus joint training is A for the FNN | +| Protocol (4.1) | 5 random splits, 315 test / 314 train geometries; RRMSE per test geometry over every node; curves averaged over the 5 splits | S. The **RRMSE normaliser is not defined: A** | +| Published targets | Table C.5 FNN(100,4,ReLU): t_fill RRMSE median **0.0339**, min 0.0130, max 0.2565. Table C.6 GBRT(100,8): median **0.0338**, min 0.0082, max 0.2550. Text: 300 of 314 test geometries well below 0.1; ML only slightly better than naive (Fig. 6, naive median not tabulated) | S | +| Code | None found in the paper; fim-python is public | M. Grok to search | + +## 2. Plan (tonight; CPU, plus a GPU only for the FNN if useful) +- **R0 (Grok, 30 min, read-only).** Is the Zenodo 10027027 record public, how large, and what fields (V-dot or flow rate, gate)? fim-python version and tetFIM API (gate as node set or faces). Any author code (GitHub/GitLab, Greif or Meyer, "eikonal injection molding"). +- **R1 (Mold, paper only, before any run).** Fix the ambiguous items with one choice each, recorded: + - V*_f measure: lumped nodal volume, strict "<", ties reported; + - nodal gradient: volume-weighted average of P1 element gradients; + - D_i at cell centres: mean of the vertex values; + - S: 6 unique components; + - RRMSE primary = RMSE / RMS(y) per geometry; also report RMSE / max(y) and RMSE / mean(y) as secondary, because the paper does not define it; + - analytic synthetic checks from Fig. 4: on a rod V*_f is linear in D_f; on a centre-gated disk it is quadratic. +- **R2 (MeltAstra, implementation).** Pinned project venv with fim-python, xgboost and scikit-learn (record the versions); feature pipeline per section 1 plus Mold's choices; synthetic tests first. + - Data: v4 (10027027) if downloadable at a reasonable size, otherwise v5 (624 parts, flagged as a version deviation). + - Target: t* = t_fill V-dot/V if V-dot is available, otherwise t_fill/t_set, flagged. +- **R3 (MeltAstra, run).** The paper protocol: 5 random 315/314 splits (seed fixed and recorded), naive + GBRT + FNN, per-geometry RRMSE over **all nodes**, and min/median/max over the 5 splits. Budget: up to 40 worker CPU-h and up to 2 GPU-h tonight; per-part guards as before; failures kept. +- **Reproduction acceptance (fixed now, before any result).** "Reproduced on the paper's protocol" means: + - GBRT and FNN fill-time median RRMSE (primary normaliser) in [0.028, 0.040], around the published 0.034; + - the naive model is within about 25 % of GBRT's median (the paper's "ML only slightly better than naive"); + - the max stays in the same regime (0.2 to 0.3). + - Anything outside these is reported as a deviation with its likely cause (data version, RRMSE normaliser, target scaling), not tuned away. +- **R4 (after R3, possibly tomorrow).** The same method on our grouped 617/209 protocol and S2 query domain. Compare against R2-G24 and the S2 B0, both at the paper's per-geometry RRMSE and at our component-mean relative L2. This answers directly how much of the R2/B0 gap was the Dijkstra approximation. + +## 3. Note on a result that matters +Gre24 reports that its learned models are **only slightly better than the naive V*_f model** on fill time. In our S2, R2 (7.30 %) is well ahead of our Dijkstra B0 (10.40 %). If the faithful eikonal naive model closes most of that gap, then our B0 was mostly a weak distance construction, and the remaining public fill-time problem is smaller than S2 suggested.