Files
MeltMind/docs/claude/planning-discussion-2026-11-19.md
T
2026-11-19 05:56:37 +00:00

46 lines
5.2 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Planning discussion (Claude hosts, 2026-11-19; requested by Yanqi)
This discussion changes no current experiment authorisation. Running: 1b fold 0 (GPU0). In preparation: Track B GINO reproduction (environment acceptance first).
Where we stand (details in `research-tree.md` and `stage-review-2026-11-19.md`):
- H2 (learned mesh representation) is our best model: field 7.08 → 6.58 % vs A on the grouped Greif protocol, confirmed; last-fill location worse; seed spread small for field, large for last fill.
- Tabular additions (B-series), last-fill losses (B17, c) and B13 context inside H2 (b) all failed to add anything.
- Greif paper comparison is not legitimate yet (our A gives 3.99 % on random splits vs the paper's 3.31-3.38 %).
- The data-sufficiency test was not identified (overfitting signs on the longer schedule; 1b tests that).
- Constraints: Moldflow student licence only (no batch generation right); Horsea 4 × V100 (GPU3 single task); fresh holdout blocked.
## Questions (answer independently, before reading anyone else's; at most one page)
1. **Short term (weeks): how do we make the current approach solid?** What is missing before H2 can be called a reliable result (seeds, folds, holdout, last fill, protocol)?
2. **Medium term: which single comparison is most worth running?** (e.g. GINO ShapeNet reproduction then H2 on it, Ahmed, Gre24 protocol reproduction, a stronger published mesh baseline on Greif, pretraining arm.)
3. **Long term: how do we return to cross-geometry injection-molding prediction?** Data (new parts, generation, licence), model, evaluation on genuinely new part families.
For each answer state: **key dependencies, cost (GPU-h, people time, licences), and what evidence would change your mind.**
## Who
MeltSol (executor + literature), MeltAssist (reviewer), MoldAstra (maths), NewMeltAstra (fresh view; has implemented nothing), TRON (coordination only, optional). Answers go to Git via MeltSol, links in Slack. Claude writes its own answer only after collecting the others (to avoid anchoring), then hosts one discussion round and rules.
## Claude's answer and ruling (after reading all four at 2c0dd614)
The four answers converge closely, so I skip a separate discussion round (saves Astra quota) and rule directly. The one difference (MeltAssist: GINO as a baseline on Greif; the others: GINO vs H2 on ShapeNet) is resolved by ordering both.
**Short term: make H2's claim exact, with zero new training.**
1. Finish 1b fold 0 (running) and read it against its pre-registered screen.
2. From existing frozen per-part scores (seed 0, seed 1, A): a family-weighted, per-seed × per-fold paired table of field error, last-1 % tail region and hard endpoint, keeping failed-part denominators. Pre-registered questions: is the field gain concentrated in a few families (leave-one-family-out sign flips)? Do both seeds miss the true tail on the *same* parts (a systematic failure worth a mechanism study) or on different ones (noise)? MeltSol builds it, MeltAssist checks it; 0 GPU-h, about 1-2 person-days.
3. H2's claim stays limited to: field gain on the grouped Greif protocol; last-fill cost; no new-family or paper-level claim. No third seed until the table shows it would change a decision.
**Medium term: one comparison chain, in this order.**
1. Track B: reproduce GINO on ShapeNet-611 (gate 6.62-7.62 %, three seeds).
2. H2 adapted to pressure (identity-labelled "H2-P"; molding-specific inputs removed) vs GINO on the same 500/111, same raw geometry, equal training budget and tuning allowance, three paired seeds, preprocessing and inference costs reported. This is our route to a public 3D benchmark claim and tests whether H2's representation carries over without molding priors.
3. Then GINO as a strong published baseline on Greif fill time (MeltAssist's proposal), same grouped protocol and inputs as H2, two folds × two seeds first. If GINO matches or beats H2 there, it becomes the default baseline.
Each step needs its own pre-registered gate and review; budgets come from measured epochs + 25 % (current estimates 17-50 GPU-h for step 2, 20-45 for step 3).
**Long term: genuinely new families, with the inputs that make fill time identifiable.**
- The learning target is fill time given geometry, gate/inlet history, material and thermal boundary (MoldAstra): data without these cannot be fixed by a bigger model.
- Evaluation design: whole-family holdout plus paired interventions inside a new family (one gate or process change per part, NewMeltAstra), reporting whether the model ranks the two designs correctly, not only field error.
- Zero-GPU first: MeltSol inventories existing archives with CAD, process conditions, solver version and reuse terms (3-5 person-days).
- **Blocking dependency: data rights.** The student licence does not cover batch generation. A 10-family pilot (10 parts × 2 conditions = 20 reference solves) needs either a licence or partner data. This is a question for Yanqi.
**What would change this plan:** GINO not reproducible (pause the ShapeNet ranking, go to step 3 directly); H2-P consistently behind GINO at equal inputs and budget (H2 stays a molding model, GINO-type becomes the 3D default); the family table shows the gain concentrated in a few families (narrow the claim before scaling anything).