Which Manufacturing Constraints to Switch On: Causal Pruning for Topology Optimization Pipelines

Engineers who design parts by computer optimization, for example topology optimization (an algorithm that decides where material goes inside a design domain to make a part stiff and light), rarely use just one manufacturability safeguard. They stack several: a filter that sets a minimum member thickness, a projection step that pushes densities toward pure solid or void, an overhang rule so the part can be 3D-printed, and a penalty against disconnected islands of material. Each safeguard has a strength setting. Every combination costs a full, expensive optimization run, so the combinations multiply quickly, and the usual default is to switch everything on. This proposal treats each safeguard as an intervention, meaning a deliberate change to the system, in a small causal graph. It then borrows tools from causal bandits to discard combinations that cannot help before any run is spent.

The core bet rests on one fact about manufacturing. Performance and fabrication feasibility are reached mostly through a handful of cheap, measurable intermediate quantities of the finished geometry: thinnest member, overhang area, number of connected components, and the fraction of in-between "gray" densities. If those quantities carry most of the effect, an engineer's rough cause-and-effect diagram can rule out most safeguard combinations, and a surrogate model that passes information through those quantities can learn from far fewer runs. The proposal also tests a sharper and riskier claim. Switching on every safeguard at once may be the manufacturing version of the "all-at-once" mistake that Structural Causal Bandits: Where to Intervene? showed can be strictly suboptimal. If so, a minimal set of safeguards should beat the full stack at equal feasibility.

Core Claim

The claim has three parts. The first is about sample efficiency. Constrained Bayesian optimization (a method that fits a probabilistic model to past runs to choose the next run) over the settings of a topology-optimization pipeline will find a configuration that is both feasible and high-performing in far fewer full solves when it uses three ingredients:

  1. Pruning by causal graph, in the style of cMISReduce from Constrained Causal Bayesian Optimization.
  2. A surrogate that routes predictions through measured geometric quantities, as Malek's graph-based multi-task Gaussian process routes them through intermediate graph nodes.
  3. A statistical check that falls back to ordinary optimization when the engineer's graph is wrong, adapted from HAC-UCB in Adaptively Exploiting d-Separators with Causal Bandits.

The second part is substantive. The best configurations will usually switch on fewer safeguards than the all-on default. The extra safeguards add optimization difficulty, such as the local minima that the continuation schedules criticized in Normalized field product approach produce, without adding feasibility. The third part is a falsifiable limit. The advantage should be largest at small evaluation budgets and should shrink as a generic method learns which settings matter.

Where This Sits Relative to Prior Work

Causal pruning so far has lived on small tabular graphs. Lee and Bareinboim's possibly-optimal minimal intervention sets, the counterfactual extension in Counterfactual Structural Causal Bandits, and the multi-objective version in Multi-Objective Causal Bayesian Optimization all assume a known graph over a few scalar levers. They were evaluated on synthetic or medical problems. Malek's cCBO is the closest method: it prunes intervention sets with no path to the objective or constraints, and it reports 100% feasibility on the HEALTH benchmark against 10–20% constraint violations for non-causal constrained optimization. We reuse its machinery almost unchanged. What is new is the domain: the levers are manufacturability mechanisms in a physics solver, and the mediators are geometric descriptors computed from the output design.

The manufacturing literature motivates the gap but solves a different problem. TopoDiff (Diffusion Models Beat GANs on Topology Optimization) and Bernoulli flow matching (On the Generalization in Topology Optimization via Sensitivity-Conditioned Bernoulli Flow Matching) build manufacturability into a learned generator. The latter uses a causal Markov chain only to justify conditioning on sensitivity fields. Neuroshaper (Shaping freeform nanophotonic devices with geometric neural parameterization) and DiffOPC (Differentiable Edge-based OPC) build rules into differentiable parameterizations. All of these still expose penalty weights, filter radii, and rule choices that someone must set.

The obvious intersection idea is to run causal pruning over pixels or edge segments. We deliberately reject it. Adjoint gradients already cost the same regardless of how many pixels there are, and continuum physics couples every region to every other, so a sparse pixel-level graph would be invented rather than real. The causal structure that genuinely exists sits one level up, between the pipeline's settings and the geometry they produce.

Causal-PIK plays a supporting role. Its kernel measures similarity between actions by their downstream effects rather than their raw coordinates, and our mediator-routed surrogate rests on the same intuition. Its own ablation, where AUCCESS falls from 65.0 to 42.0 when the effect-based kernel is replaced with a Euclidean one, is the strongest evidence in the file that this intuition pays off.

Research Program

We will build the causal design graph. Its levers are the minimum-length-scale filter radius, the projection sharpness and its schedule, the overhang filter angle and on/off status, the connectivity penalty weight, the volume fraction, and the build direction. That is roughly eight to ten levers, discretized to three levels each, giving several thousand combinations. The mediators are thinnest member thickness from morphological opening, overhang area fraction at the build direction, connected-component count, and grayness. The outcomes are compliance and a binary feasibility verdict from independent rule checkers.

Edges from levers to mediators largely follow from construction. The overhang filter changes overhang area, for instance. We will also include direct lever-to-compliance edges, because a safeguard can hurt performance by making optimization harder, not only by changing geometry. The graph is written down by the engineer and is not learned. This matches the finding in Graph Learning is Suboptimal in Causal Bandits and Causal Bandits with Unknown Graph Structure that recovering the full structure is unnecessary and costly.

The optimizer then runs in four steps:

  1. cMISReduce removes lever subsets that have no directed path to any active constraint or to compliance.
  2. Malek's surrogate predicts the mediators from the levers and then predicts compliance and feasibility from the mediators.
  3. Constrained expected improvement chooses the next run.
  4. A misspecification test, run every few evaluations, compares the surrogate's predictions with those of an unstructured Gaussian process on held-out runs. If the causal model is rejected, the optimizer switches to unstructured constrained optimization over the full lever space.

We accept the cost that HAC-UCB's impossibility result (its Theorem 6.2) makes unavoidable. No method can achieve both the best rate under a correct graph and the best worst-case rate. We will report both regimes rather than claim both.

Proposed Experiments

The tasks come from the file. They are the cantilever, MBB beam, and compliant displacement-inverter problems in 2D and 3D, as used for the normalized field product method. We add a sample of boundary-condition instances from the 33,000-structure SIMP dataset used by TopoDiff, run here through a classical solver rather than a generator. We also add a small set of CFD-TO channel-flow problems, to check that the graph transfers to a second physics.

Each evaluation is one complete optimization run. Budgets are 10, 20, 30, 50, and 100 runs, averaged over 20 seeds. The metrics are:

  • the best compliance among feasible designs at each budget, normalized by an unconstrained SIMP reference;
  • the number of runs needed to reach the first feasible design within 5% of the best design found by exhaustive search on 2D instances, where exhaustive search is affordable;
  • the fraction of evaluated runs that come out infeasible, a measure of wasted fabrication-relevant compute.

Baselines

The naive reference point is the all-on default. Every safeguard is enabled at textbook strengths from the normalized-field-product and SIMP literature and run once per task. Beating it means one of two things. Either a selected configuration matches its feasibility rate with lower compliance, which is the minimal-set claim, or it reaches feasibility where all-on fails. The file contains no reported all-on result for these benchmarks, so we measure it ourselves. The nearest published figures are the normalized field product method's grayness below 1% and TopoDiff's roughly 6% floating-material rate on out-of-distribution conditions. They describe different pipelines and serve only as sanity anchors.

A second simple baseline uses the normalized field product method alone, which embeds a length scale without tuning, plus only the overhang rule. Its purpose is to test whether a parameter-free method already removes most of the search problem.

The strongest competitor is standard non-causal constrained Bayesian optimization. It uses independent Gaussian processes per outcome, with automatic relevance determination (each lever gets its own learned length scale, so the model can discover that a lever does not matter) and constrained expected improvement. It runs over the full lever space with the same budgets, the same initial designs, the same acquisition optimizer, and the same hyperparameter-fitting routine. The only reported number for this family is Malek's: 10–20% constraint violations on HEALTH against 100% feasibility for the causal surrogate. No reported result was found on Lacuna for constrained Bayesian optimization over topology-optimization pipeline settings. Beating it means a statistically significant gain in feasible compliance at budgets of 30 runs or fewer, together with a lower infeasible-run fraction.

Three ablations isolate the ingredients. The first applies pruning but uses unstructured surrogates. The second uses the mediator-routed surrogate without pruning. The third runs the full method with a deliberately corrupted graph, in which one true edge is deleted and a false mediator is asserted, both with and without the misspecification test.

Why not just…?

Why not just let automatic relevance determination discover which levers matter? This is the most serious threat. With enough runs, learned length scales will flag inert levers without any graph. The causal method's edge must therefore show up at small budgets, which is exactly where 3D solves make budgets small. Its other potential advantage is the mediator layer, which shares information between feasibility and compliance in a way per-lever length scales cannot. The budget sweep settles the question. If the gap closes by 20 runs, we will report that causal pruning is a cold-start aid rather than a general improvement.

Why not just switch everything on, or use a parameter-free method? Practitioners do exactly this, and the normalized field product method was designed to end manual tuning. However, it addresses length scale only. It does not address overhang or build direction, and the file documents that continuation schedules can trap designs in poor local minima. If all-on matches the best selected configuration on most tasks, the minimal-set claim fails. The proposal then reduces to faster tuning, and we will say so.

Why not just make every rule a differentiable penalty and optimize once, as Neuroshaper and DiffOPC do? Those methods do produce rule-clean designs natively. Even so, they carry penalty weights and feature thresholds whose values trade efficiency for feasibility: Neuroshaper's 95.0% against 96.2% for an unfabricable pixel design is one such trade. Those weights are precisely our levers, so the approaches are complementary rather than competing. A follow-up would apply the same optimizer to Neuroshaper's penalty weights on the mode-converter problem.

Decisive Checks

The project succeeds if three things hold. Causal constrained optimization must beat automatic-relevance-determination optimization on feasible compliance at 10 to 30 runs across both 2D and 3D tasks. The selected configurations must use fewer safeguards than all-on while matching its feasibility. The corrupted-graph run with the misspecification test must stay within a modest margin of the unstructured baseline, while the same run without the test degrades. If the first check fails, the causal graph adds nothing over generic relevance learning in this domain, and that negative result is worth publishing. If only the second fails, the contribution narrows to sample efficiency.