From 6eaa3aa4dba8c24426213a75ba6f4033cdb8648c Mon Sep 17 00:00:00 2001 From: Claude Date: Fri, 20 Nov 2026 16:04:20 +0000 Subject: [PATCH] next-stage design options draft Co-Authored-By: Claude Opus 5.5 --- docs/claude/design-options-2026-11-20.md | 33 ++++++++++++++++++++++++ 1 file changed, 33 insertions(+) create mode 100644 docs/claude/design-options-2026-11-20.md diff --git a/docs/claude/design-options-2026-11-20.md b/docs/claude/design-options-2026-11-20.md new file mode 100644 index 000000000..d3f3f997c --- /dev/null +++ b/docs/claude/design-options-2026-11-20.md @@ -0,0 +1,33 @@ +# MeltMind next-stage design options (Claude draft, 2026-11-20 16:10 UTC; for Yanqi's design discussion) + +Status labels: [verified] with evidence in research-tree.md; [inferred]; [untested]. MoldSol and MeltSol will refine; this is the skeleton. + +## 0. What the product is +A surrogate for Moldflow filling: from part geometry, gate location(s) and process conditions known before the run, predict the fill-time (arrival-time, seconds) field over the whole 3D part. [goal] + +## 1. What we can already claim [verified] +- Field error on 98 unseen public families: A 7.87 % vs H2 6.55 / 6.50 % (same process, single gate). +- Last-fill region: k4 readout improves H2 (+3.5 pp, CI [+0.6, +6.2], 320 / 132 clean Greif parts); vs A on shared nodes the difference is uncertain. +- Not tested: varying gates, varying process, multiple gates, design choices. + +## 2. Option A: data (needed for anything beyond single-process) +- Source: Yanqi's Moldflow licence (product/version pending). Per-node fill time is documented for Insight/Synergy (XML/Patran/ASCII, Python `moldflow` API: PlotManager.get_scalar_data, export_to_vtk) [verified docs, Grok]. Adviser: documented route ends at MFR for viewing [verified docs]; MeltWin is checking a real MFR read-only [pending]. +- Pilot: 10 new parts from new families × 2-3 gate positions × 2-3 process settings (flow rate / injection time, melt temperature), same material; mesh-sensitivity check on a subset (2 meshes) to measure solver variability [untested; cost to be measured on the first part]. +- Stop: if per-node export with node IDs or coordinates cannot be verified on one part, or if mesh variability exceeds the differences between options. + +## 3. Option B: model +- H2 extended with process inputs (flow rate, melt/mould temperature) and gate encoding that supports several gates (distance-from-each-gate channels or a learned gate embedding) [design]. +- Keep the physics-anchored residual on the arrival-time clock; keep k4 readout for last fill. +- Comparison arm: GINO with the same inputs (running/pending) to choose the backbone. + +## 4. Option C: evaluation (what "useful in design" means) +- Field error (seconds and relative) and spatial overlap of the filled region at shared physical times. Total filled volume alone is not enough: two equal-volume branches filled in the wrong order give the same volume curve (TRON). Summing the per-time spatial difference over a window equals volume-weighted truncated arrival-time L1, so time cuts are for interpretation, not an extra score. +- Last-fill region hit (k4) and distance. +- Design ranking: for each part, do model and Moldflow order the gate/process options the same way (e.g. which option fills a named thin rib to 95 % first, or which balances two branches)? Ties within Moldflow's own mesh variability are 'unjudgeable'. +- Report separately: new geometry / seen process; seen geometry / new process; both new (TRON). +- Literature: Lee et al. 2026 (GAT, six process inputs) is limited to one product family; no published method selects families × gates × process settings to train a cross-family field surrogate [Grok, read]. + +## 5. Decisions for Yanqi +1. Licence product/version (determines data route). +2. First engineering decision to support (thin-rib fill order, branch balance, short shot, pressure limit) and the tolerance that changes a decision. +3. Pilot size and which part families (her own parts or public ABC geometries).