Optimization Technology¶
A concept note for the AI for Automotive Moulded Parts Engineering study. Referenced by the vendor reports wherever a vendor labels classical optimization as "generative AI." This is explanatory background (textbook material), not new cited research.
Several vendors in this study market products under names like "Generative Design" or "Autonomous Optimization." The technology underneath is usually topology optimization or design-of-experiments — mathematical methods that long predate modern AI. This note explains what those methods actually do, and why the distinction from generative AI matters.
Topology optimization — what it actually does¶
Topology optimization is a numerical method from the late 1980s (the founding paper is Bendsøe & Kikuchi, 1988). It works like this:
The setup. You give the software three things:
- A design space — the maximum volume the part is allowed to occupy (e.g., "anything inside this bracket envelope").
- Loads and constraints — where forces are applied, where the part bolts to something, and which regions must stay untouched (bolt holes, mating faces).
- An objective — usually "maximum stiffness" or "minimum weight," plus limits such as "stress must stay below X."
The loop. The software then runs a repeated cycle:
- Fill the design space with material and run a finite-element analysis (a physics simulation) to see how stress flows through it.
- Identify which regions carry significant load and which sit nearly idle.
- Remove material from the idle regions; keep or reinforce the loaded ones.
- Re-run the simulation on the new shape. Repeat — typically tens to hundreds of iterations — until the shape stops improving.
The result is the distinctive organic, bone-like structure these tools are known for: material left only along the load paths. It looks creative, which is part of why it markets so well as "generative."
Why it is not generative AI¶
The key difference is where the answer comes from:
| Topology optimization | Generative AI | |
|---|---|---|
| Source of the answer | Physics equations, solved on your problem, every time | Patterns learned from training data |
| Learns from past parts? | No — carries nothing between jobs | Yes — that is its entire mechanism |
| Same inputs twice? | Same answer (deterministic) | Typically varies |
| Improves with use? | No | Can, if retrained on new data |
Topology optimization is an algorithm executing physics — the same family of methods as mould-flow simulation, just pointed at shape instead of plastic flow. A generative AI model that had seen thousands of brackets could propose a new bracket by analogy, without solving the physics from scratch. These are fundamentally different mechanisms, even when the output looks similar.
Does this apply to moulded parts?¶
Yes — but with a twist that is itself instructive.
Raw topology-optimization output is usually unmouldable. The organic, bone-like shapes violate the core injection-moulding DFM rules almost by construction: they are full of undercuts (the part cannot be ejected from a two-half mould), they mix thick and thin sections (sink marks, voids, warpage, long cooling — breaking the uniform-wall-thickness rule), and they carry no draft angles. Such shapes suit 3D printing or machining, not a mould.
So for moulded parts the method is used in two adapted ways:
- With manufacturing constraints bolted on. Serious tools offer a demould / draw-direction constraint: the optimizer may only form shapes that can be pulled out of the mould along its opening direction (no undercuts), plus minimum/maximum member-size limits standing in for wall-thickness rules. Dassault's own SIMULIA Tosca Structure (from the FE-Design acquisition, 2013) supports exactly this kind of demould control.
- As a rib-layout guide — the everyday use in automotive plastics. Take a door-panel carrier or a bracket, fix the base wall, and let the optimizer show where the load paths run. A designer then reinterprets the fuzzy result into proper mouldable ribs, applying the textbook rib-to-wall thickness ratio (roughly 50–60% of the wall, to avoid sink marks) and draft by hand. The optimizer proposes; the moulding rules are applied afterwards, by a human.
Note what this means for the study's argument: even inside "generative design," all the actual moulding knowledge — draft, rib ratios, sink-mark avoidance, ejection — enters as classical rules and constraints, not as anything learned from data. The AI label sits on top; the moulding intelligence underneath is exactly the rules-based kind.
The related case: design-of-experiments (DoE)¶
The same relabelling appears with "autonomous optimization" claims (e.g. SIGMASOFT's "Autonomous Optimization" for moulding process set-up). The machinery there is virtual design-of-experiments: run many simulations across a grid of candidate process settings (gate positions, temperatures, timings), then pick the best performer. It is automated search, powered by brute-force simulation — again deterministic, again learning nothing from data.
Why the AI label goes uncontested¶
If the label is "just sitting" on classical technology, one might expect someone to challenge it. Nobody does, because every force in the market pushes the same way:
- "AI" has no policed definition — and vendors have a defensible position. No regulator or trademark owns the word. Academically, AI has always included search, optimization and rule-based expert systems — the "good old-fashioned AI" of 1970s–80s textbooks. A vendor calling topology optimization "AI" is not strictly lying; they are using the 40-year-old academic definition while letting the audience hear the modern meaning (learning systems, generative models). Misleading by connotation is very hard to contest.
- No damaged party. False-advertising claims need someone who was harmed. But the tool genuinely works — the optimized part really is lighter and stiffer. A product that works as advertised, mislabelled only as to mechanism, injures nobody in a way a court or regulator cares about.
- Everyone at the table benefits from the label. The vendor gets valuation and relevance; the buyer's executives get to report AI adoption to their board (many companies have literal "adopt AI" mandates, and buying "generative design" ticks the box); the trade press gets an AI headline. And competitors will not contest it because they all wear the same label on similar technology — calling out a rival's AI-washing invites scrutiny of your own. Mutual disarmament.
- The experts who can tell the difference don't care to fight. Simulation engineers mostly know it is the same optimization they have used for twenty years. The label is a marketing layer above their workflow; contesting it wins them nothing.
- Where enforcement exists, it is aimed elsewhere. Regulators have begun acting on "AI-washing" — the US SEC charged two investment advisers over false AI claims in 2024, and the FTC ran "Operation AI Comply" — but those target consumer and investor deception. B2B engineering-software claims are technical, the audience small and expert, the materiality low: the last place enforcement will reach.
The consequence for this study: the market will not police the label, so the study must. That is why every such case is tagged [Marketing] and held against the test below — and why the honest negative findings in the vendor reports (e.g. "no AI reaches the moulding domain") appear nowhere in vendor material and had to be established by deliberate absence-of-evidence searching.
The honest test used throughout this study¶
Does the system learn from data, or does it compute from equations?
Only the first is AI in a meaningful sense. Both are genuinely useful — topology optimization and virtual DoE are mature, valuable engineering tools — but calling them AI borrows a halo they did not earn. The vendor reports flag every such case as [Evidence + Marketing] and cross-reference this note.