8.0 KiB
MeltMind research proposal v2 (Claude, 2026-11-13)
Supersedes v1 (1PYTHLxnE3YdOKIfuHXNDqILeVs9_mZKu). It incorporates Grok's precedents, MoldAstra's two-example framing, TRON's statistics correction and Yanqi's clarified intent.
0. Intent (Yanqi, 2026-11-13)
Large-model surrogates are being explored across fluid mechanics. Injection-molding filling is a plausible place for a practically useful surrogate to arrive first: low Re, a quasi-static pressure solve, and a structured thin-walled geometry. The aim is solid, frontier-level work that could later feed the team's molding software. Geometry organisation is a means to cross-part generalisation, not a goal.
Both main candidates get a full, careful attempt. Neither is cut early to save effort, and neither is written up as a proven advantage.
1. Why these two candidates: lineage, not improvisation
O2: local canonical coordinates (thin-wall local frames)
Lineage in our own work. The original "thin-wall geometry unification" study:
- derived the pull-back of the thin-gap pressure problem: A_ref = sqrt(det g) h^3 g^-1 / (12 mu), capacity density sqrt(det g) h (math core D.2);
- compared multi-atlas reference blocks vs a common background domain vs complexes (approach page, 2026-10-27);
- showed that invertible transforms do not remove complexity by themselves (E.1), so any benefit must be measured.
My first-round proposal in this thread (reference fluidity log S_ref) and Frank's correction (at fixed mu it is affine-equivalent to log h) point the same way: the useful content is the local thin-gap structure, not a single scalar.
Precedents.
- Global deformation to a reference domain: Geo-FNO (Li22), DIMON (Yin24b), Bur23. These need one diffeomorphism per part and are limited to one topology family (Yin24b Discussion; Li22 §3.4-3.5).
- O2 is the local-atlas version: frames are built per region from d(x) (normal from grad d, in-plane directions, thickness-normalised through-gap coordinate), so no global map and no topology constraint.
Potential.
- Recurring wall patches become nearly identical inputs across parts and poses.
- The thin-gap physics structure (through-thickness coordinate, metric) enters the input explicitly.
- Directly addresses Type A generalisation (same local geometry, different pose).
Main risk. Frames are ill-defined at junctions, rib roots, bosses and thick regions, and frame ambiguity (sign, in-plane rotation) must be handled.
Competing routes for the same goal.
- E(3)-equivariant architectures (equivariance built in instead of canonicalisation).
- Data augmentation with rotations/reflections.
- Global maps (only within one topology).
O1: grouping/aggregation aligned with geometric type
Lineage.
- Yanqi's 2026-11-11 idea of cutting by local geometry type.
- The block-basis discussion in this thread (Jia26b, Wu25; Frank's two paper criteria: re-cutting the same part gives the same answer, reconnecting gives the correct different response).
- The agreement to keep full block fields, shared local operators and an explicit global path.
Precedents.
- RIGNO (region aggregation + data-efficiency experiments): the closest threat.
- Huang SNI and NEST: local operators + Schwarz.
- Transolver slices: a learned soft partition.
- None is 3D filling, and none tests type-aligned grouping against a matched generic partition.
Potential. Shares statistics and computation among recurring structures (walls, ribs), much as CAE engineers reason by features. It may help at small training budgets.
Main risk. Learned global slices may already capture it; and a gain may come from partitioning per se, not from types.
Competing routes. Learned partitions (slices), generic graph partitions, multi-scale graphs.
2. Other routes kept on the map (not dropped; not first)
- R1 whole-part frontier models at scale (Transolver++/GINO-type) with physics features. This is our strong reference in every comparison.
- R2 feature engineering + classical ML (Gre24 eikonal/geodesic + GBRT/FNN on the same corpus). A mandatory, cost-accounted baseline; may be strong for fill time.
- R3 hybrid physics: a fast in-house 2.5D/Hele-Shaw solve plus learned 3D correction. Possibly the most practical for the team's software, since the solver is owned. Its natural home is the in-house stage, with same-part 2.5D/3D pairs.
- R4 explicit global coupling (coarse pressure/mass solve inside the model), needed for Type B generalisation (reconnection) and for pressure targets.
3. What "generalisation" means (MoldAstra's framing)
- Type A: same local geometry, different pose. Tests coordinate handling. Testable on public data without new labels: rigidly rotating/translating/mirroring a test part together with its gate leaves its fill-time field unchanged. Scaling is NOT invariant.
- Type B: same local geometry, different connection (e.g. an upstream bypass). Tests whether real connectivity is learned. Not causally testable on v5; in-house stage.
- Type C: unseen part families (source-group / shape-cluster holdout). The main public-data target.
Every result states which type it addresses.
4. Hypotheses (unchanged in substance)
- H-A: input becomes more familiar under organisation (model-free; ranks operators, never refutes them).
- H-B: sample efficiency on unseen source groups.
- H-C: mechanism by region.
- H-D: robustness to a strong encoder and to re-cutting/re-meshing.
5. Roadmap: two parallel tracks on shared infrastructure
Shared (S0, running):
- data contract;
- static data card;
- d(x) validated;
- near-duplicate rule;
- frozen split;
- baselines R2 and B0;
- one strong reference model (R1 family) with identical inputs;
- power estimate from baseline error spread across groups.
Track O2 (local frames):
- MoldAstra: frame definition, invariances, ambiguity handling, failure set.
- MeltAstra: implementation and failure-rate census.
- S1: H-A coverage + Type A rotation diagnostic.
- S2: learning curves, O2 vs the same model without frames (and vs augmentation as the competing route).
Track O1 (type grouping):
- MoldAstra: type definition from validated descriptors.
- MeltAstra: matched partitions with acceptance tests.
- S1: H-A coverage + affine diagnostic D.
- S2: learning curves vs a matched generic partition and a control cut.
Joint S3, only if both show positive effects: 2x2 factorial (frames x grouping), mechanism by region, strong-encoder check, re-cut stability.
S4, in-house CAE:
- Type B reconnection tests;
- gate/process/material interventions;
- pressure-limited short shots;
- R3 hybrid route;
- product screening metrics.
6. Decision rules (TRON's correction)
Δ = relative error reduction vs the track's controls; [L, U] = 95 % interval; m = preregistered margin (candidate 10 %).
| Region | Outcome |
|---|---|
| L > m | meaningful gain supported |
| 0 < L <= m and U >= m | positive effect, size undetermined |
| 0 < L and U < m | positive but below margin |
| L <= 0 and U < m | no meaningful gain |
| L <= 0 and U >= m | insufficient evidence |
| U < 0 | harmful |
- Point estimates are reported but never decide.
- Single-seed runs are exploratory.
- With about 35-40 test groups, "positive, size undetermined" is realistic. The power estimate is done before GPU planning.
7. Resources
- S0-S1: Horsea CPUs.
- S2: 4x V100 16 GB, one run per GPU. The two tracks can run in parallel; both share the reference-model runs.
- All GPU-hours come from a measured pilot (t_step, memory) using the formula in v1 §8. No numbers before that.
8. Ownership
- Claude: design, integration, acceptance criteria, review.
- MoldAstra: mathematics (H-A metric, frame and type definitions, analytic checks, independent review).
- MeltAstra: Horsea implementation and runs.
- Grok: SoA survey of fluid foundation models and molding surrogates, precedent and novelty challenges, written sources.
- TRON: relay.
9. Not claimed
- That O1 and O2 are the best of all possible routes.
- Any advantage before S2 results.
- Anything about gate/process/material or short shots from public data.