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

[Copilot]

The issue asks which base unit-cell structures and design variables should drive the Bayesian optimization campaign for TPU + PETG multi-material 3D-printed tensegrity-inspired energy absorbers — note PETG, not the PLA currently described in proposal.tex.

The repo is LaTeX-only (no BO code), so this is a research/scoping task. Following the established Edison-driven research pattern in this repo, a non-blocking literature query was submitted, the completed result is now committed under edison-trajectories/, and a follow-up synthesis cross-references the design-gap families surfaced by PR #22 and recasts the combined set of axes as an Ax HierarchicalSearchSpace.

Changes

• No source/proposal edits — deferred until PETG-vs-PLA scope is confirmed.
• Edison LITERATURE_HIGH task — task_id 5ae24eaf-5b6e-45cf-9f6c-1c7fbd881738 submitted, fetched (status: success), and committed:
  • edison-trajectories/tpu-petg-bo-variables-5ae24eaf-5b6e-45cf-9f6c-1c7fbd881738.md — verbatim formatted_answer (Question + cited Answer + numbered References).
  • edison-trajectories/tpu-petg-bo-variables-5ae24eaf-5b6e-45cf-9f6c-1c7fbd881738.json — full structured model_dump_json for reproducibility.
• PR Add reference STL models for canonical tensegrity structures + Edison literature surveys (initial + design-gaps follow-up) #22 design-gaps synthesis — edison-trajectories/tpu-petg-bo-variables-additions-from-pr22.md cross-references the design-gap families from the PR Add reference STL models for canonical tensegrity structures + Edison literature surveys (initial + design-gaps follow-up) #22 Edison survey (6226a551-…: Snelson X-module, Pajunen truncated-octa, Oster reentrant auxetic, Rhode-Barbarigos pentagonal ring, Pugh diamond / zig-zag, tensegrity torus, Hanaor double-layer grid, bistable double-prism, 6-bar wheel patent, etc.) against the original 5ae24eaf variable table and identifies the additional BO axes those families introduce:
  • 8 new categorical axes: extended unit_cell list, chirality_pattern, cable_routing (Pugh diamond/zig-zag/circuit), tensegrity_class (1/2/k), tendon_clustering, bistable_mode, payload_carrier, tessellation_pattern.
  • 5 new or conditional continuous axes: per-tendon prestress group fractions, N_bay stack count, re-entrancy angle (auxetic family), ring radius / strut-circuit count (ring & torus families), hub/cradle mass fraction.
  • Constraint: defer class-2 / class-k topologies until the issue Extensively explore and validate the joint design for PETG + TPU prints to mimic tensegrity #38 Phase-4 joint designs validate strut-to-strut contact; reaffirms TPU 85A (not 95A) as the lab tendon material.
• Hierarchical search space (§D of the additions doc) — per facebook/Ax#140, the combined parameter set is recast as an Ax HierarchicalSearchSpace:
  • Root topology_family ChoiceParameter with 8 branches: prism_stack, icosa_class, trunc_octa, ring_torus, auxetic_periodic, double_layer_grid, bistable_cell, x_module_weave.
  • Per-branch dependents={...} map demoting chirality_pattern, cable_routing, payload_carrier/hub_mass_fraction, tessellation_pattern, bistable_mode, tendon_clustering, re_entrancy_angle, ring_radius/N_circuit, and the per-tendon prestress simplex from flat axes to conditional children of the relevant branch.
  • joint_validated_class2 fixed feature on the Ax Experiment to gate the ring_torus / class-2 branch on issue Extensively explore and validate the joint design for PETG + TPU prints to mimic tensegrity #38 Phase-4 joint validation.
  • Surrogate / acquisition guidance: stratified Sobol seed (≥3 specimens per family) before BO turn-on, with a MultiTaskGP alternative (topology family as task feature) for the ~50–100-specimen budget.

Query scope (original 5ae24eaf task)

• (A) Base unit-cell topologies — 3-/4-bar prisms, octahedron, icosahedron, expanded octahedron, truncated tetrahedron, Pajunen-style tensegrity-inspired cells, stacked-prism / lattice tilings; manufacturability and reported energy-absorption notes.
• (B) Design variables to expose to BO:
  • Geometric/topological: strut L, D, slenderness L/D, cable/skin cross-section, prestress, twist, tiling Nx·Ny·Nz, relative density, connectivity.
  • Material/print: PETG layer height, infill % & pattern; TPU shore (85A/95A), wall count, infill; nozzle/bed temps; speed; PETG–TPU wrap thickness (Ye 2023, Khatri 2024).
  • Loading: drop height, impactor mass, strain rate — fixed, not BO vars.
• (C) PETG-vs-PLA implications for sensible bounds (toughness, lower stiffness, lower Tg → strut slenderness, infill, layer-adhesion ranges).
• (D) Recommended numeric bounds + categories sized for a BoTorch/Ax search space and a ~50–100-specimen undergrad campaign.
• (E) Primary objectives (peak transmitted force, SEA, compaction efficiency, plateau stress) and constraints (mass, print time, cycle durability).

Follow-up

• When the BO search space is first encoded, consume the original 5ae24eaf table plus the PR Add reference STL models for canonical tensegrity structures + Edison literature surveys (initial + design-gaps follow-up) #22 synthesis additions using the §D hierarchical tree — concretely as Ax ChoiceParameter(dependents=...) Python or JSON — and apply the joint_validated_class2 fixed feature for the class-2/k deferral constraint.
• If PETG is confirmed as the strut material, update proposal.tex parameterization wording (currently PLA+TPU) accordingly.

[sgbaird]

@copilot fetch and commit results

[Copilot]

@copilot fetch and commit results

Done in 3478122 — committed both edison-trajectories/tpu-petg-bo-variables-5ae24eaf-5b6e-45cf-9f6c-1c7fbd881738.md (formatted answer with citations) and .json (full structured task) for task 5ae24eaf.

[sgbaird-yolo]

Worth seeing if there are additional parameters to consider based on new results in #22 ; still running there. Can consider later

[sgbaird-alt]

@copilot go for it in #24 (comment)

[Copilot]

@copilot go for it in #24 (comment)

Done in 39a2e7c — synthesised the PR #22 design-gaps survey (6226a551-…) against the original 5ae24eaf TPU+PETG variable list and committed the delta as edison-trajectories/tpu-petg-bo-variables-additions-from-pr22.md. Eight new categorical axes (X-module / auxetic / bistable / payload-carrier / tessellation-pattern / chirality / cable-routing / class / tendon-clustering), five new or conditional continuous axes (per-tendon prestress fractions, N_bay, re-entrancy angle, ring radius/circuit count, hub mass fraction), plus a constraint to defer class-2/k topologies until issue #38 Phase-4 joint designs validate strut-to-strut contact.

[sgbaird-alt]

@copilot I'm guessing this probably needs to be treated as a hierarchical search space (see facebook/Ax#140 for context)

[Copilot]

@copilot I'm guessing this probably needs to be treated as a hierarchical search space (see facebook/Ax#140 for context)

Agreed — recast as an Ax HierarchicalSearchSpace in 89a2bd9. Added §D to edison-trajectories/tpu-petg-bo-variables-additions-from-pr22.md with: (i) an 8-branch topology_family root categorical (prism_stack / icosa_class / trunc_octa / ring_torus / auxetic_periodic / double_layer_grid / bistable_cell / x_module_weave) and the per-branch child parameters, so axes like re_entrancy_angle, ring_radius, payload_carrier/hub_mass_fraction, cable_routing, bistable_mode, tendon_clustering, per-tendon prestress simplex, etc. are demoted from flat to conditional ChoiceParameter(dependents=…) children per facebook/Ax#140; (ii) the issue #38 Phase-4 class-2 gate as a joint_validated_class2 fixed feature that hides the ring_torus branch until joints are validated; (iii) a stratified-Sobol seed (≥3 specimens per family) and a MultiTaskGP alternative for the ~50–100-specimen budget.

[sgbaird]

comment from @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

@copilot thoughts? Send a high effort edison literature query related to this. Fetch this session

[Copilot]

comment from @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 thinkin...

Done in b3e49d6 — submitted + fetched Edison LITERATURE_HIGH task 5191cf4d-873a-4e3e-9077-9565a2602ba1 in this session and committed edison-trajectories/heterogeneous-params/{heterogeneous-params-5191cf4d-…md,…json,README.md} plus the idempotent driver scripts/edison/submit_heterogeneous_params.py.

Thoughts (echoing Edison §a-g):

The literature backs @me-madsen's intuition for some parameters but pushes back on a naive per-member expansion of all of them on PR #35:

1. Precedent exists, but converges to orbits, not per-member. Skelton/Nagase 2014, Goyal/Skelton/Peraza-Hernandez 2020, Xu 2018, Zhang 2021, Pajunen 2019 all treat per-member cross-section / prestress as LP/MINLP variables — and in every case the optimum clusters into a small number of discrete groups corresponding to symmetry orbits (saddle / top / bottom triangles + struts for a T3-prism), not a continuum. Pajunen 2019 specifically gets a 3.1× SEA gain at +3.6% mass by changing the strut-vs-cable diameter ratio — a per-orbit, not per-member, knob.
2. Symmetry: T3-prism has D3. All 3 struts live in one orbit; the 9 cables decompose into 3 orbits of 3 (saddle/top/bottom). The honest dimensionality of "per-member diameters" is therefore 4 orbit diameters, not 12. Breaking that symmetry breaks the integral self-stress state (Chen 2018; Masic 2005) and shifts the equilibrium geometry — fine if intentional (Micheletti 2013 shows it can unlock bistability), but expensive to explore on a 50–100 specimen budget.
3. FFF on Bambu H2D quantizes hard. Tuncel 2024: PETG cross-sections deviate ~11.5%, practical diameter steps ~0.6–0.8 mm on a 0.4 mm nozzle. TPU 85A bridging penalizes mid-print diameter transitions. Edison recommends treating cable diameters as a 4–5-bin categorical ({1.2, 1.8, 2.4, 3.0, 4.5} mm) and strut diameters in ~1 mm steps — anything finer is below the noise floor.
4. High-dim BO: SAASBO first, then TuRBO. SAASBO (Eriksson & Jankowiak 2021) is the strong fit for "most members don't matter, a few do" sparse-effect regimes up to ~25-D; TuRBO + qNEHVI (Eriksson 2019; Daulton 2020) for the O(100+)-D tiling regime. Both have first-class Ax/BoTorch hooks. Sits cleanly under the existing PR Explore TPU+PLA BO design variables and base structures (Edison literature result + PR #22 design-gaps synthesis, hierarchical search space) #24 §D HierarchicalSearchSpace (per facebook/Ax#140) — per-member axes nest as children of the topology_family choice.

Concrete recommendation for the PR #35 follow-on batch:

Parameter	Current (PR #35)	Recommended (per Edison §f)
R_mm, H_mm, twist_deg, infill %	scalar ✓	keep scalar
strut_d_mm	1 scalar	1 strut_orbit_d_mm ∈ [3.5, 9.0] continuous (D3 symmetry: all 3 struts same diameter)
cable_d_mm	1 scalar	3 cable_orbit_d_mm (saddle / top / bottom) ∈ {1.2, 1.8, 2.4, 3.0, 4.5} categorical
(new) per-cable prestress	—	9-element simplex (sum = 1) — strongest case for fully-per-member expansion (Nagase & Skelton 2014 minimal-mass)
BO engine	Sobol-only (1 batch)	SAASBO + qNEHVI, stratified Sobol seed ≥3/family, MultiTaskGP across sim ladder (PR #33)

Net: take diameter dims from 2 → 4 (per-orbit), add a 9-D prestress simplex if the joint design supports differential prestress per cable, and turn on SAASBO instead of pure Sobol after the first batch. Avoid going fully-per-member on diameters until orbit symmetry is exhausted.

Full cited write-up + Ax/BoTorch recipe in edison-trajectories/heterogeneous-params/heterogeneous-params-5191cf4d-873a-4e3e-9077-9565a2602ba1.md.