Generative engineering design & topology optimisation

Design engineers traditionally draw a part, simulate it, and iterate by hand. Topology optimisation and generative design invert this: you state loads, constraints, material and manufacturing process, and software proposes many geometries that meet them with minimal mass or maximal stiffness. Newer learned models (diffusion, neural fields, surrogates) make exploration faster and let engineers steer it in natural language. Typical results are lighter parts, fewer physical prototypes and shorter design cycles, always verified by classical simulation.

Problem & value

• Today: parts are designed from experience and incrementally improved; weight, cost and performance are traded off by few manual iterations. Classical topology optimisation (SIMP, level-set) exists since the 1990s but yields organic shapes that need manual CAD clean-up and can be slow for large 3D models.
• AI value: broader exploration of design alternatives, manufacturing-aware constraints (additive, casting, 2.5D milling), multi-objective ranking, faster iteration through learned surrogates, and lower barrier for non-specialists. Motivation is often material and energy savings (lightweighting in automotive, aerospace, machinery, tooling, fixtures).
• Benefits: vendors and case studies report substantial mass reductions and fewer part counts through consolidation; figures vary strongly by part and baseline, so measure on your own reference parts rather than rely on marketing numbers [1][2].

Variants

• Classical topology optimisation inside CAD/CAE: density-based, level-set, lattice infill.
• Generative design in cloud CAD: multi-process, multi-material candidate sets with cost estimates.
• Implicit/field-driven design for additive manufacturing: lattices, graded structures, heat exchangers.
• Learned generative models: diffusion or GAN-style models conditioned on loads and volume fraction, trained on optimisation results; also generative models for 2D/3D shape, airfoil or mechanism design [3].
• LLM-driven design agents: natural-language problem set-up that calls a solver through tools (see SCI-024 for text-to-CAD).
• Private/maker use: hobby 3D printing brackets via Fusion personal licence or open-source tools.

How it works

1. requirements & loads (PLM/requirements tool, SCI-013)
2. design space + keep-out zones (CAD)
3. problem set-up (loads, supports, materials, process constraints, objectives)
4. generation: topology optimisation / generative solver / learned generative model
5. candidate filtering and Pareto ranking
6. geometry reconstruction (smooth mesh
7. B-rep/CAD)
8. verification FEA/CFD (independent, high-fidelity solver)
9. manufacturability check (DfAM/DfM, cost estimate PRED-032)
10. human review, test and release via PLM

• Optimisation core: gradient-based solvers (finite elements with adjoint sensitivities) remain the workhorse; evolutionary and Bayesian methods handle non-differentiable or multi-physics goals.
• Learned components: (a) surrogates predicting stress or compliance to speed loops (SCI-005); (b) conditional generative models that propose near-optimal layouts in a forward pass, later refined by a solver; (c) VLM/LLM interfaces for set-up and reporting.
• Human-in-the-loop: engineers define the problem, select among candidates, and own sign-off. Load cases and constraints are the most error-prone inputs; a wrong boundary condition gives a confidently wrong optimum.

Data preparation

Data needed: design spaces and interfaces (CAD/STEP), load cases and boundary conditions, material data, manufacturing constraints. In an SMB these live in CAD/PDM systems, simulation reports on file shares, and often in the heads of senior engineers or in Excel load tables.

Typical gaps: load cases undocumented or only as PDF reports, inconsistent materials, missing keep-out zones, models with unclean geometry, and no record of why past designs were chosen.

Preparation steps: 1. Write down standard load cases, materials and manufacturing rules per part family (PREP-016, PREP-015). 2. Clean CAD (defeature, watertight STEP) and fix naming in PDM (PREP-010). 3. Collect past designs with verified FEA results as benchmarks (PREP-018). 4. Store material and process parameters as tidy tables (PREP-011).

Tools: CAD defeaturing functions, PDM metadata, a simple parameter spreadsheet or database.

Share of effort: about 10-20 percent; classical topology optimisation needs little historical data, learned models need more.

Implementation paths

Which path for whom

• SMB: Path A, because mature add-ons to CAD/CAE suites (Fusion, Inspire, nTop, Creo) exist at SMB prices and the process is standard.
• Enterprise: Path B, because generative design must be wired into PLM, in-house load cases and cloud HPC for many FE solves.
• Developer: Path C, because open solvers and optimisation libraries allow custom objectives and manufacturing constraints.

Path A — Buy (recommended for SMBs)

• Integrated CAD/CAE: Autodesk Fusion (generative design), Siemens NX and Simcenter, Dassault Systemes (CATIA, SIMULIA, 3DEXPERIENCE; EU), PTC Creo Generative Topology Optimization, Altair Inspire / OptiStruct, Ansys Mechanical topology optimisation, Hexagon MSC (Apex/Nastran).
• Implicit/AM-focused: nTop (implicit modelling), Materialise (EU) software, Oqton, Dyndrite.
• EU vendors: Dassault Systemes (FR), Siemens (DE), Altair (acquired by Siemens) and Ansys (acquired by Synopsys); check current status.
• Choose A when: you already license a CAD/CAE suite; the add-on is the cheapest route.

Path B — Build on cloud APIs

• Cloud CAE/HPC (AWS, Azure, Google; EU regions) running licensed or open solvers; orchestrate with Python (e.g. OpenMDAO, pyoptsparse) and expose through an internal tool.
• Frontier LLMs (Claude, Gemini, GPT, Mistral) via function calling to translate requirements text into solver parameters and to draft reports; the numerical work stays in deterministic solvers.
• Choose B when: you need custom objectives, batch sweeps or integration with PLM.

Path C — Self-host (open source)

• Topology optimisation: FEniCSx / dolfin-adjoint, Firedrake, MFEM, TopOpt codes, ToPy, PyTopo3D, Calculix and Code_Aster (EDF, FR) as FE solvers, OpenMDAO for workflows.
• Learned generative models: PyTorch with diffusion or neural-operator libraries (NVIDIA PhysicsNeMo, DeepXDE); public datasets from topology-optimisation papers [3].
• Mesh/CAD: Gmsh, Open CASCADE (build123d, CadQuery), FreeCAD.
• Hardware: HW2 for 2D and small 3D; HW3 for large 3D models and training surrogates.

Path D — Custom training

Worth it when you repeatedly optimise one part family (e.g. brackets, wheel rims, heat sinks): generate thousands of solver-optimised designs, train a conditional generative model or surrogate, and use it for instant first proposals. Needs hundreds to thousands of solved cases and ML expertise; always keep a solver in the loop for final verification.

State of the art (2026)

• Commercial tools combine topology optimisation with manufacturing constraints and automated CAD reconstruction; "generative design" in mainstream CAD is routine for lightweight brackets, housings and tooling.
• Research shows diffusion models and neural operators reproduce optimiser outputs in seconds and support conditioning on loads and volume fraction; surveys note that physical validity, 3D scaling and constraint satisfaction are not guaranteed [3].
• LLM agents increasingly set up and run CAE workflows, but reliability on boundary conditions and units is still limited; treat them as assistants, not decision-makers.
• Limits and failure modes: unconstrained optimum ignores fatigue, buckling, impact, thermal cycling or inspection access; additive parts have anisotropy and residual stress; learned generators may output geometries that violate constraints or are not manufacturable; reproducibility of results across tools is limited; reconstructed CAD can differ from the optimised mesh.

Shared resources & combinations

• Simulation core shared with SCI-005 (surrogates), SCI-022 (test-data analytics) and SCI-016-style optimisation workflows; GPU HW3 serves both training and inference.
• Upstream: SCI-013 (requirements/MBSE), KNOW-020 (reuse of existing parts before designing new).
• Downstream: SCI-023 (CAM/NC programming), VIS-057 (monitoring of AM builds), PRED-032 (should-cost).
• Interface: SCI-024 (text/sketch-to-CAD) for natural-language entry and drawings.

Effort, cost & KPIs

• Path A: weeks for first parts if licences exist; main cost is licences and engineer training.
• Path B: 2-4 months for an integrated pipeline; HPC cost driven by number of FE solves.
• Path C: 3-6 months; requires FE and optimisation expertise.
• Path D: 6+ months incl. dataset generation.
• KPIs: mass or cost reduction versus baseline design at equal verified performance, number of design iterations, time from requirement to released design, share of candidates passing manufacturability checks, prototype and test count, material waste.

People & change

• Involve early: management, the design department lead, design and FEA engineers (end users), IT for licences and compute, data owners of the PDM/CAD archive. Manufacturing planners should join to confirm constraints.
• Works council: depends. No evaluation of people is intended, but CAD/cloud tools log user activity, which can trigger §87(1) no. 6 BetrVG (objective suitability suffices); §90(1) no. 3 requires information when work processes change. See GUIDE-004. Public sector: Personalrat.
• Data protection: essentially no personal data; a short DPO check and an AVV with cloud vendors suffice.
• AI literacy: under AI Act Art. 4 (since 27 Jul 2026) employer and provider must take measures to support staff AI literacy; plan a short training (see GUIDE-006).
• Pilot and communication: pick one bracket or housing and a small team; compare against the hand-designed baseline and share results openly (GUIDE-008, GUIDE-009).
• Typical concerns: loss of design craft and organic shapes that are hard to make; address by keeping manufacturability checks and designers as final authors.
• Human control: engineers define goals and constraints, review candidates and release only designs verified by classical FEA.

Risks, regulation & compliance

• EU AI Act: typically minimal risk; design tools are not Annex III. Safety components of products fall under sector law (Machinery Regulation (EU) 2023/1230, Medical Device Regulation, aviation EASA rules); the AI design tool itself usually is not the regulated product but its output is. AI literacy duty (Art. 4) applies.
• Product liability and certification: the new Product Liability Directive (EU) 2024/2853 covers software; keep traceable verification evidence (verification FEA, tests, review sign-off) for critical parts. Certified parts (aerospace, medical implants, pressure equipment) need qualified processes irrespective of how geometry was generated.
• GDPR: generally not applicable; engineering data only (personal data limited to user accounts and author metadata).
• IP and trade secrets: designs and training sets are sensitive; prefer EU hosting or on-prem for cloud CAE, check vendor terms on training with customer data; export-control checks for dual-use designs.
• Mitigations: independent verification solver, physical testing for critical parts, design reviews, version control of set-up assumptions.

Getting started

1. Pick 2-3 reference parts with known weight and test data; record baseline mass, cost and iterations.
2. Standardise load-case and constraint templates with your structural engineers.
3. Run the add-on in your existing CAD/CAE suite (Path A) and validate against your baseline.
4. Check manufacturability and cost with your production partner or AM supplier.
5. Only then consider surrogates or learned generators for part families you optimise repeatedly.

Sources

1. Autodesk — Generative design in Fusion — https://www.autodesk.com/solutions/generative-design
2. Altair — Inspire topology optimisation — https://altair.com/inspire
3. Regenwetter, Nobari, Ahmed — Deep Generative Models in Engineering Design: A Review (J. Mechanical Design, 2022) — https://arxiv.org/abs/2110.10863
4. nTop — implicit modelling platform — https://www.ntop.com
5. FEniCSx project — https://fenicsproject.org
6. Code_Aster (EDF) — https://www.code-aster.org
7. OpenMDAO — https://openmdao.org
8. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
9. Directive (EU) 2024/2853 (Product Liability) — https://eur-lex.europa.eu/eli/dir/2024/2853/oj