Explore TPU+PLA BO design variables and base structures (Edison literature result + PR #22 design-gaps synthesis, hierarchical search space)

edison-trajectories/heterogeneous-params/README.md

@@ -0,0 +1,60 @@
+# Heterogeneous (per-member) BO parameter axes — Edison literature trajectory
+
+Source: PR #24 comment
+[4520542433](https://github.com/vertical-cloud-lab/tensegrity-optimization/pull/24#issuecomment-4520542433)
+(sgbaird relaying @me-madsen):
+
+> noting that we could also allow for diameters of individual struts/cables
+> to vary, rather than assuming a fixed diameter. Similar for other
+> parameters perhaps. Mostly thinking in context of #35 right now
+
+PR #35 currently sweeps a single `strut_d_mm` and a single `cable_d_mm` per
+T3-prism specimen (all 3 struts and all 9 cables tied to one diameter each).
+The proposal is to expand selected scalar parameters into vector / per-member
+parameters — taking the per-specimen design vector from ~5-D to O(12-30+) D.
+
+## Trajectory
+
+| File | Notes |
+| --- | --- |
+| `heterogeneous-params-5191cf4d-873a-4e3e-9077-9565a2602ba1.md` | Verbatim `formatted_answer` from Edison `LITERATURE_HIGH` task `5191cf4d-873a-4e3e-9077-9565a2602ba1` (job: `LITERATURE_HIGH`, status: `success`). 466 lines covering (a) precedent in Skelton/Nagase/Goyal/Zegard/Pajunen/Xu/Zhang minimal-mass tensegrity sizing, (b) form-finding / bistability / energy-absorption implications under D3 symmetry, (c) FFF manufacturability on Bambu H2D, (d) high-dim BO methodology (SAASBO, TuRBO, ALEBO, additive GPs, hierarchical search spaces, equivariant GPs, MTGP), (e) symmetry exploitation under C3/D3, (f) numeric recommendations + Ax/BoTorch recipe, (g) ranked failure modes, (h) numbered references. |
+| `heterogeneous-params-5191cf4d-873a-4e3e-9077-9565a2602ba1.json` | Full structured `model_dump_json` for reproducibility. |
+| `../../scripts/edison/submit_heterogeneous_params.py` | Idempotent submission + polling driver (reads `EDISON_PLATFORM_API_KEY` or mirrors `EDISON_API_KEY`; uses `client.get_task(task_id=...)` polling; writes a `-SUBMITTED.json` placeholder so a follow-up session can resume on wall-clock timeout). |
+
+## Headline recommendation (Edison §f)
+
+For the first heterogeneous PR #35 follow-on batch on a single T3-prism:
+
+- **Keep scalar:** `R_mm`, `H_mm`, `twist_deg`, infill %.
+- **Expand to per-orbit (not per-member)** the strut and cable diameters,
+  exploiting the T3-prism `C3 → D3` symmetry that decomposes the 9 cables
+  into 3 orbits (saddle / top / bottom triangles): one `strut_orbit_d_mm`
+  continuous axis ∈ [3.5, 9.0] mm + three `cable_orbit_d_mm` categorical
+  axes ∈ {1.2, 1.8, 2.4, 3.0, 4.5} mm (FFF-resolvable bins on a 0.4 mm
+  nozzle, per Tuncel 2024 dimensional-accuracy data).
+- **Expand to fully per-member** only the per-cable prestress fraction
+  simplex (sum = 1), per Skelton's minimal-mass prestress optimization
+  (Nagase & Skelton 2014; Goyal, Skelton & Peraza Hernandez 2020).
+- **BO engine:** SAASBO for ≤25-D campaigns (Eriksson & Jankowiak 2021),
+  escalate to TuRBO + qNEHVI for tilings that push into O(100+) D
+  (Eriksson 2019; Daulton 2020). Initialize with stratified Sobol (≥3
+  specimens per topology family / orbit configuration).
+- **Symmetry handling:** hard-enforce orbit symmetry as the default
+  search space (Approach (i) in §e); only relax to per-member axes if
+  the symmetric Pareto front is exhausted, since the lab budget is
+  50–100 specimens.
+
+## Cross-references
+
+- PR #24 hierarchical search space synthesis:
+  [`../tpu-petg-bo-variables-additions-from-pr22.md`](../tpu-petg-bo-variables-additions-from-pr22.md)
+  §D recasts the topology axes as an Ax `HierarchicalSearchSpace` per
+  [facebook/Ax#140](https://github.com/facebook/Ax/issues/140). This per-member
+  expansion nests one level *below* that hierarchy — inside any chosen
+  `topology_family` branch, selected scalar parameters expand into
+  per-orbit / per-member vector parameters.
+- PR #35 (`bo/t3_prism_sobol_batch.py`) is the concrete target for the
+  Edison §f recipe.
+- PR #33 sim-ladder (MuJoCo → Newton/Warp XPBD → PolyFEM+IPC / DiffPD) is
+  the recommended host for the MTGP / multifidelity escalation discussed
+  in Edison §d.