Atomizer/hq/workspaces/shared/war-room-isogrid/review-optimizer.md

War-Room Review: Adaptive Isogrid Plate Lightweighting — Optimizer Perspective

Reviewer: Optimizer Agent
Date: 2026-02-19
Spec Reviewed: technical-spec.md (Architecture Locked)

---

1. STRENGTHS

• Clean separation of concerns. The Python Brain / NX Hands split is architecturally sound. JSON-only geometry transfer eliminates format conversion drift and makes the pipeline deterministic and debuggable.

• Reserved-region strategy is the right call. Preserving load/BC topology by only replacing sandbox geometry avoids the single worst failure mode in optimization-in-the-loop FEA: broken references causing silent solve failures or garbage results.

• Manufacturing constraints enforced at generation time. This is correct. Checking manufacturability post-hoc and penalizing is far more expensive and creates flat, uninformative penalty landscapes. Rejecting bad geometry before FEA saves ~90 seconds per infeasible trial.

• 15 parameters is tractable for TPE. The parameter count is within Optuna TPE's effective range. The space is all-continuous, which TPE handles well.

• Gmsh Frontal-Delaunay is a good mesher choice for this class of problem. Background size fields map naturally to the density field formulation.

---

2. WEAKNESSES

2.1 Objective Function Design — Critical Issues

The single-objective penalty formulation is the weakest part of this spec.

• Penalty weight of 1e4 is arbitrary. The mass objective is in grams/kg scale. The penalty 1e4 * ((σ/σ_allow) - 1)² creates a landscape where a 10% stress violation contributes ~100 to the objective. But what's the mass scale? If the plate is 500g, the optimizer will happily trade 500g of mass for a tiny stress violation reduction. If the plate is 50g, the penalty dominates everything. The penalty scaling is not normalized to the problem.

• Quadratic penalty creates flat regions. When stress is well below allowable, the penalty gradient is zero — the optimizer gets no signal about stress margin. When stress is far above allowable, the quadratic explodes and the optimizer can't distinguish between "terrible" and "slightly less terrible." An augmented Lagrangian or log-barrier would be more informative.

• return float('inf') for invalid geometries is poison for TPE. TPE builds kernel density estimators from the trial distribution. Inf values create degenerate distributions. Use a large finite penalty instead (e.g., worst observed + margin), or better yet, use Optuna's pruning/fail mechanism (raise optuna.TrialPruned()).

2.2 Parameter Space Risks

• s_min and s_max can cross. Nothing in the spec prevents s_min > s_max being sampled. This would invert the density-to-spacing mapping. Need either a constraint (s_max > s_min + gap) or reparameterization (e.g., sample s_min and s_range = s_max - s_min).

• t_min and t_0 can conflict. If t_0 < t_min, the clamp in the thickness formula makes t_min the effective thickness everywhere regardless of density. This creates a flat region in parameter space where t_0 and gamma have zero effect — wasting exploration budget.

• R_0 and R_edge are in absolute mm. For plates of different sizes (200mm vs 600mm), fixed bounds [10, 100] mm have completely different meaning. These should be normalized to plate characteristic length or at minimum the bounds should be plate-size-adaptive.

• Several parameters are likely redundant/coupled. eta_0 + alpha + beta together control the DC offset and scale of the density field. There are likely degenerate combinations (e.g., high eta_0 + low alpha ≈ medium eta_0 + medium alpha for uniform-ish fields). This wastes TPE's budget exploring equivalent solutions.

2.3 Convergence Concerns

• 2 min/iteration × 500-2000 trials is 17-67 hours. This is fine for a production run but terrible for development iteration. The spec has no surrogate-assisted strategy to reduce evaluation count.

• No warm-starting. If you change the plate geometry (different project), the optimization starts from scratch. No transfer learning or prior injection.

• No early termination per trial. If the Python Brain produces a geometry with 3 pockets (clearly degenerate), we still run a full 90-second FEA. The spec should define cheap pre-solve rejection criteria beyond just valid.

---

3. OPTIMIZATION OPPORTUNITIES

3.1 Multi-Objective Formulation (High Priority)

This should be Pareto, not penalty-weighted. The problem is naturally bi-objective: minimize mass, minimize max stress (or maximize stress margin). Optuna supports NSGA-II/MOTPE natively via optuna.create_study(directions=["minimize", "minimize"]).

Benefits:

• Eliminates arbitrary penalty weights entirely
• Produces a Pareto front the engineer can select from
• Reveals the mass-stress tradeoff structure — is it convex? Are there cliff edges?
• Allows adding displacement as a third objective without reformulating

3.2 Dimensionality Reduction (High Priority)

15 parameters with heavy coupling is inefficient. Concrete reductions:

1. Reparameterize spacing: Replace s_min, s_max with s_center, s_range (or just s_ratio = s_min/s_max + s_max). Eliminates the crossing issue and reduces effective DOF.

2. Fix manufacturing params early. r_f, d_keep, w_frame, t_min are manufacturing constraints, not design variables. Fix them from manufacturing input, optimize only the 10-11 remaining design parameters. Run a sensitivity analysis first to confirm they're low-impact.

3. Merge decay exponents. Using one p for both hole and edge influence is already done. Good. But R_0 + kappa + R_edge is 3 parameters controlling influence radii. Consider: is the edge term even necessary? The perimeter frame (w_frame) already handles edge reinforcement structurally. beta and R_edge may be redundant with it.

3.3 Surrogate-Assisted Optimization (Medium Priority)

At 2 min/eval, a Gaussian Process or Random Forest surrogate could:

• Pre-screen candidate trials (reject predicted-bad before FEA)
• Provide gradient estimates for local refinement
• Optuna's BoTorchSampler wraps GP-based BO and handles 15D reasonably

However: the geometry validity check (pre-FEA) already acts as a cheap filter. The real question is whether the FEA landscape is smooth enough for a surrogate to learn. With a density-field parameterization, it likely is — small parameter changes produce small geometry changes produce small stress changes. Recommend trying BoTorch after 100-200 TPE trials as a refinement phase.

3.4 Sensitivity Analysis (High Priority, Pre-Optimization)

Before running 500 trials blind:

• Run a Sobol or Morris screening (50-100 cheap evaluations) to identify which parameters actually matter
• Likely finding: alpha, R_0, s_min, s_max, t_0 dominate; kappa, gamma, p are second-order
• Fix or narrow low-sensitivity parameters → effective dimension drops to 6-8 → TPE converges 3-5× faster

---

4. PIVOT CONSIDERATIONS

4.1 Is TPE the Right Sampler?

For v1, yes. But not as the only sampler.

TPE is a reasonable default for 15D black-box with moderate budget. However:

• TPE struggles with parameter interactions (it models marginals independently). The density field has strong alpha×R_0 and s_min×t_0 interactions.
• CMA-ES would likely outperform TPE here once you're past initial exploration. The objective landscape (mass as a function of smooth density field parameters) is likely quasi-convex and continuous — exactly CMA-ES's strength.
• Recommended: Hybrid strategy. 100 trials TPE (exploration) → switch to CMA-ES (exploitation). Optuna supports this via CmaEsSampler with warm-start from TPE trials.

4.2 Should We Use Topology Optimization Instead?

No, and the spec gets this right. Here's why:

• SIMP/level-set topology optimization produces organic, freeform material distributions. The output needs interpretation/smoothing to become manufacturable isogrid. You'd optimize → interpret → re-validate, adding a lossy translation step.
• The density-field-parameterized isogrid is directly manufacturable by construction. Every trial output is a valid waterjet/CNC geometry. This is a massive advantage.
• Topology optimization has its own high-dimensional space (element-wise densities, thousands of variables) requiring adjoint solvers. The current approach uses 15 parameters with a commercial solver. Much simpler to implement and debug.

However: If v1 results show the isogrid pattern is consistently suboptimal compared to what topology optimization would suggest (e.g., the optimal rib pattern isn't triangular at all), then a hybrid approach — use SIMP to identify load paths, then fit an isogrid to those paths — would be worth investigating for v2.

4.3 Is the Density Field Formulation Sound?

Mostly, with one fundamental concern.

The density field is driven purely by geometric proximity to holes and edges, weighted by user-assigned importance. This assumes that structural importance correlates with hole proximity. For many bracket/plate problems, this is reasonable — holes are load introduction points.

But it fails when:

• Load paths don't pass through high-weight holes (e.g., a lightly-loaded mounting hole near a high-stress bending region)
• The critical stress region is in the middle of the plate, far from all holes and edges
• Multiple load cases create competing stress patterns

The v2 stress feedback addresses this, but v1 is flying blind on actual structural behavior. This is the biggest technical risk: v1 may converge to a geometrically elegant but structurally mediocre design.

Mitigation: Compare v1 optimum against uniform isogrid early. If v1 isn't meaningfully better, the density field hypothesis is wrong and stress feedback (or a different parameterization) is needed sooner.

---

5. WHAT'S MISSING

5.1 Multi-Objective Pareto Front

Already discussed in §3.1. This is the single highest-impact improvement. A penalty-weighted single objective hides the problem structure.

5.2 Robustness / Uncertainty Analysis

• No mention of manufacturing tolerances. If the optimal rib thickness is 2.1mm and t_min is 2.0mm, a 0.2mm waterjet kerf variation kills it.
• No mention of material property uncertainty.
• At minimum: after finding the optimum, perturb each parameter ±5-10% and verify the design doesn't collapse. Better: add a robustness term to the objective (e.g., optimize worst-case over a small parameter ball).

5.3 Multiple Load Cases

The spec assumes a single solve setup. Real plates have multiple load cases (operational, limit, fatigue). The objective should aggregate across load cases (e.g., max stress across all cases, mass is load-case-independent).

5.4 Buckling

Shell plates with pockets can buckle. The spec mentions "extensible to modal/buckling" but doesn't include it in v1 objectives. For thin plates (6mm thickness, 60mm cell size), pocket buckling is a real failure mode. At minimum, add a buckling eigenvalue check as a constraint.

5.5 Fatigue / Stress Concentration

Pocket corners are stress risers. The fillet radius r_f directly affects fatigue life. The spec treats r_f as a free optimization parameter, but it should have a hard lower bound driven by fatigue requirements, not just manufacturing.

5.6 Baseline Comparison

No mention of comparing against: (a) solid plate (upper mass bound), (b) uniform isogrid (simpler approach), (c) existing hand-designed lightweighting. Without baselines, you can't quantify the value of adaptive density.

5.7 Parameter Coupling Diagnostics

No plan to measure or visualize parameter correlations, interaction effects, or identify flat/degenerate regions in the search space during the optimization. Optuna's built-in visualization (parameter importances, contour plots) should be explicitly planned.

---

6. RECOMMENDED DIRECTION

Immediate (Before First Run)

1. Switch to multi-objective (mass, max_stress) using Optuna MOTPE or NSGA-II. Drop the penalty formulation.
2. Fix the parameter crossing issue (s_min < s_max, t_min ≤ t_0). Reparameterize or add constraints.
3. Replace float('inf') with raise TrialPruned() or large finite penalties for failed geometries.
4. Fix manufacturing parameters (r_f, d_keep, w_frame, t_min) from engineering input. Reduce to ~10 free parameters.
5. Add cheap pre-FEA rejection — if mass estimate from geometry alone exceeds solid plate mass or is below physical minimum, skip FEA.

Short-Term (First 200 Trials)

6. Run Sobol sensitivity analysis (first 50-100 trials) to identify dominant parameters.
7. Compare against uniform isogrid baseline to validate the adaptive density hypothesis.
8. Add buckling eigenvalue as a constraint or third objective.

Medium-Term (After Initial Results)

9. Switch to CMA-ES for exploitation after TPE exploration phase.
10. Evaluate BoTorch surrogate for expensive-evaluation reduction.
11. Accelerate v2 stress feedback if v1 adaptive density doesn't significantly beat uniform isogrid.

The Bottom Line

The architecture is solid. The density field concept is reasonable for v1. The parameter space is tractable. The objective function formulation is the critical weakness — it needs to be multi-objective, and the penalty scaling will cause real convergence problems if left as-is. Fix the objective, fix the parameter degeneracies, and this is a viable optimization campaign. Don't fix them, and you'll burn 67 hours of compute to converge on an artifact of your penalty weights.

---

Review complete. No cheerleading detected.