Design Software History: From Structural Optimization to Generative Design: The Evolution of Topology Optimization, Geometry Kernels, and Industrial Workflows

From structural optimization to topology optimization: the academic roots

Early ideas

When structural optimization first entered engineering discourse, the central question was deceptively simple: how should material be arranged to achieve maximum stiffness with minimum weight? In 1904, A.G.M. Michell offered a mathematically incisive answer for trusses under idealized loads, deriving continuum-like truss layouts that foreshadowed what later generations would call material distribution as a design variable. Michell’s optimality conditions, framed in the language of stress trajectories and orthogonal networks, crystallized the idea that structure is a field problem, not merely a sizing chart. Decades later, as finite element analysis (FEA) matured in the 1960s–1980s through the work of pioneers like Ray W. Clough, J.H. Argyris, and O.C. Zienkiewicz, optimization seeped into the pipeline, initially as sizing optimization of bar and shell thicknesses and then as shape optimization altering boundaries with spline or nodal coordinates.

Practical algorithms filled the gap between theory and computation. Gradient-based methods such as sequential quadratic programming (SQP) and early optimality criteria approaches (popularized by researchers including L. Berke and R.T. Haftka) enabled compliance minimization and stress-constrained sizing. However, shape optimization hit walls: parametrizations could be brittle, sensitivity analysis across moving boundaries was delicate, and topological changes—like holes appearing or vanishing—were effectively disallowed. The missing ingredient was a formulation that let the topology itself become a degree of freedom. Researchers eyed the continuum, where material could be treated as a density field, but the mathematical and numerical infrastructure to do so robustly had not yet coalesced. The stage was set for a leap from boundary tweaking to interior reallocation—a transition that would define topology optimization for the next generation of tools and thinkers.

Breakthrough formulations

The late 1980s provided the inflection point. In 1988–1989, Martin Philip Bendsøe and Noboru Kikuchi established the homogenization-based framework in which microstructures (e.g., perforated composites) approximate intermediate densities, allowing continuous optimization of a pseudo-density field subject to elasticity laws. This theoretical advance lit the path for what Bendsøe, Ole Sigmund, and collaborators would popularize as SIMP (solid isotropic material with penalization): treat each finite element’s stiffness as E(ρ)=ρ^pE₀, with penalization exponent p>1 to bias toward crisp 0–1 designs. SIMP’s simplicity, paired with FEA, made large-scale topology optimization (TO) viable on industrial meshes. Meanwhile, Krister Svanberg introduced MMA (method of moving asymptotes) and its variant CONLIN in the 1990s—robust, scalable optimizers that remain default choices for thousands to millions of design variables. In parallel, Y.M. Xie and G.P. Steven’s ESO/BESO heuristics offered an appealing practitioner’s mental model: iteratively remove (and later re-add) inefficient elements based on sensitivity, producing strikingly organic load paths.

Not everyone was satisfied with gray-scale SIMP fields. In the 2000s, level-set methods—rooted in the Osher–Sethian Hamilton–Jacobi machinery—were adapted to TO by Grégoire Allaire, François Jouve, and Anca-Maria Toader, offering crisp boundary evolution without intermediate densities. Level sets excelled at curvature control and geometric regularity, linking directly to shape derivatives rather than density sensitivities. Each camp cultivated strengths: SIMP delivered ease, robustness, and turnkey multiphysics extensions; level sets delivered elegance in boundary representation and natural perimeter control. Across both, academic contributions made the discipline practical: adjoint sensitivity analysis reduced the cost of gradients to roughly one or two solves per objective/constraint, and continuation strategies (e.g., penalization p ramping and move-limits) stabilized convergence on production meshes. TO, once a theoretical curiosity, became a computational craft with repeatable recipes, rigorous mathematics, and recognizable visual signatures.

• Homogenization and SIMP: density fields, penalization, projection to drive 0–1 designs.
• MMA/CONLIN: globally reliable, sparsity-friendly optimizers with per-variable move-limits.
• ESO/BESO: intuitive add/remove heuristics that align with engineering intuition.
• Level-set TO: boundary-focused evolution via Hamilton–Jacobi PDEs and shape derivatives.
• Adjoint sensitivities: scalable gradients for large FEA systems and many load cases.

Usability and dissemination

Even the best algorithms wither without usable exemplars. Ole Sigmund’s famed MATLAB codes—“99-line” (2001) and “88-line” (2011, with Erik Andreassen and colleagues)—compressed the SIMP pipeline into digestible scripts that students, researchers, and industry engineers could read line-by-line. Those compact codes seeded a generation of PhDs and pilot projects, demystifying implementation details from sensitivity filtering to penalization ramps. As adoption widened, two failures demanded cures: checkerboarding (oscillatory 0–1 patterns encouraged by numerical instabilities) and mesh dependence (artificial detail growth on refined meshes). The community responded with filtering and regularization: sensitivity filters (e.g., Sigmund’s cone filter), density filters, and Heaviside projection schemes that sharpen designs while enforcing a minimum feature size. Perimeter control and total-variation regularization became common to suppress ragged boundaries.

Another hard lesson arrived from the factory floor: manufacturing robustness. Designs sensitive to ±tolerances or AM-induced porosity could fail qualification. Researchers including Sigmund, Søren O. Kristensen, and Brian S. Lazarov advanced robust formulations that optimize simultaneously against eroded, nominal, and dilated geometry—effectively safeguarding performance against uncertainties in fabrication. Length-scale control matured to bidirectional radius constraints; draw and symmetry constraints migrated from shape to density frameworks; and support-aware constraints for AM began to appear. TO thus crossed from elegant computation into production-worthiness, with explicit levers for minimum thickness, overhang, and post-processing allowances. The academic roots were now anchored in code, math, and pragmatics—three pillars that would support the coming industrial wave.

Industrialization: from CAE options to designer-facing tools (mid-1990s–2015)

CAE-embedded topology optimization

By the late 1990s, commercial CAE vendors started weaving TO directly into workflows that structural analysts already trusted. Altair’s OptiStruct became the banner product for density-based optimization on large industrial models, consolidating SIMP + MMA as a reliable recipe. Analysts working on automotive and aerospace structures could launch compliance-minimizing runs or shape/sizing refinements inside familiar solvers and post-processors. FE-Design’s TOSCA Structure (later acquired by Dassault Systèmes in 2013 and folded into SIMULIA/Abaqus) broadened the scope beyond compliance to handle frequency, buckling, and nonlinear objectives/constraints with a solver-agnostic coupling, turning TO into a domain-agnostic optimization overlay. Ansys integrated topology, shape, and sizing into Mechanical, and later introduced interactive exploration in Discovery—tightening the loop from idea to verified mesh. MSC Software (now part of Hexagon, acquired in 2017) embedded optimizers in Nastran workflows and eventually launched MSC Apex Generative Design to accelerate concept iteration on mid-complexity parts.

Siemens likewise embedded TO in NX/Simcenter 3D and connected it with HEEDS (via the Red Cedar Technology acquisition in 2013), formalizing a design-space exploration stack around solver-driven optimization. Across vendors, the key was not just exposing a button but stabilizing a chain: meshing that respects length scales, adjoint derivatives for dozens of load cases, globalization heuristics that avoid false minima, and post-processing capable of turning “blobby” fields into manufacturable geometry. Commercial TO proved itself by living within large-scale lifecycle tasks such as durability, NVH, and crash—in short, inside the heavy CAE where analysts spend their days. The payoff was not only performance; it was confidence. Result reproducibility across releases, clean interfaces to PLM, and scripting hooks for batch runs transformed TO from a research add-on into a standard CAE feature set.

Additive manufacturing as a catalyst

While TO matured within CAE, additive manufacturing (AM) reshaped the incentives. Powder-bed fusion, directed energy deposition, and binder jetting invited geometries that traditional subtractive constraints discouraged. TO’s characteristic trabecular networks, hole-riddled webs, and organic stiffeners now looked not merely acceptable but desirable—provided they respected AM process constraints. Vendors raced to incorporate overhang limits, minimum strut thickness, escape holes, and orientation-dependent support cost into the optimization loop. Lattice and infill research, once niche, found revenue-bearing applications as companies sought internal cooling, vibration damping, and weight reduction without tooling complexity. Autodesk’s 2014 acquisition of Within Labs signaled the strategic value of lattice and topology-aware AM kernels; Materialise’s 3-matic, long a stalwart in lattice and mesh editing, became an indispensable companion for field-based thickening, build prep, and support optimization. Machine OEMs like EOS and GE Additive amplified the message, publishing qualification guidance and parameter windows that informed optimizer-side assumptions about surface roughness, feature resolution, and fatigue knockdowns.

• Design freedoms: internal channels, graded lattices, and topology-derived webs.
• Constraints: overhang angles, minimum thickness, support minimization, powder removal.
• Property linkages: as-built surface quality, anisotropy, and heat-treatment effects.
• Economic signals: support cost, build time, nesting density, and downstream machining.

The synthesis of TO and AM reframed optimization as process-aware design. Objectives and constraints began to encode not just stiffness and mass but also printability metrics and post-processing allowances. That shift prepared the ground for a broader reframing: what if the designer’s role moved from defining a single model to defining goals, materials, processes, and constraints—and letting software produce a portfolio of viable options?

From analysts to designers

Mid-2010s tool strategies converged on one insight: analysts need precise control, but designers need guidance and speed. Altair’s solidThinking Inspire (later Altair Inspire) distilled OptiStruct’s capabilities into a workflow that non-specialists could wield: define loads and supports on a sculpted envelope, specify manufacturing methods, and receive optimization-driven proposals that were tightly coupled with interpretation tools. Ansys lowered the barrier with Discovery’s real-time physics approximations, enabling interactive feedback during modeling. MSC’s Apex Generative Design focused on usability and GPU acceleration to reframe waiting time as exploration time. Siemens paired NX’s modeling with Simcenter 3D solvers and HEEDS-driven parameter sweeps so multi-run optimization fit inside a designer’s daily cadence. Across the board, the language shifted from compliance-only targets to multi-criteria dashboards—stiffness, mass, safety factor, and manufacturability dials that stakeholders could weigh.

Importantly, PLM hooks matured. Results could be stored as configurations, linked to requirements, and formally reviewed. That governance mattered: organizations needed lineage between inputs (load cases, constraints), solver versions, and post-processing steps to justify sign-off and certification pathways. The net effect was a cultural one: topology optimization was no longer a post-rationalization of a designer’s idea; it became the first sketch—an algorithmically generated hypothesis that teams could assess, refine, and validate. This set the stage for the next semantic leap: generative design, where “optimal” ceded to “option sets,” and the UX moved decisively toward curation rather than a single answer.

Repurposing topology optimization into “generative design”: workflows, kernels, and UX (2015–present)

The UX reframing

Generative design reframed the question from “What is the optimum?” to “What is the set of viable options under my goals and constraints?” Instead of a single density field to be thresholded, users expect portfolios of candidates ranked by multi-objective metrics: stiffness-to-weight, displacement limits, cost, and factor of safety—often across multiple load cases and materials. The designer-in-the-loop specifies not just forces and fixtures but also manufacturing methods (3-axis milling, 2.5-axis, die casting, sheet metal forming, or AM with overhang limits), along with material classes and stock sizes. Candidate generation and ranking become an interactive process: adjust allowable drilling directions, set draft angles, define minimum fillet radii, and the engine recomputes filtered design spaces that respect those decisions.

In this UX, constraints become first-class sliders rather than buried solver flags. Preferences are expressed as trade-offs rather than absolutes—“accept 10% more mass to gain 20% lower cost” or “sacrifice 3% stiffness for a 90% reduction in support volume.” Visualization pivots to transparency and traceability: per-candidate heat maps of expected deflection, bar charts of cost drivers (stock, machining, or build time), and flags for DFM rule breaks. Generative tools also distinguish between exploration and commitment: a candidate’s geometry must be editable using familiar operations, or at least convertible into a stable format that downstream CAD, CAM, and CAE can consume without losing fidelity. The overarching UX principle is curation: the human sets intent, the software casts a net across feasible design space, and together they converge on a manufacturable, cost-aware, and certifiable choice.

Platform shifts and compute strategies

The compute model that supports portfolios is necessarily parallel. Autodesk’s Fusion 360 Generative Design, born from Project Dreamcatcher within Autodesk Research under leaders including Erin Bradner and Mike Haley, embraced cloud-scale batch solving, where dozens or hundreds of candidates run concurrently with elastic resource allocation. That scale changes not just speed but behavior: exploration is quantified, and dead-ends are cheap. In desktop environments, vendors invested in GPU acceleration—Ansys Discovery, MSC Apex Generative Design, and GPU-enabled pipelines from Altair and Siemens—so that meshing, sensitivity evaluation, and projection steps could iterate at interactive cadence. Hybrid strategies emerged: lightweight on-device screening and preconditioning, with heavy solves dispatched to clusters or cloud nodes.

• Elastic compute: spin up many solves per load case and material, prune dominated options early.
• GPU-native meshing and solvers: accelerate candidate “sprint cycles” for near-real-time feedback.
• Incremental updates: warm-start from prior candidates when sliders change marginally.
• Telemetry and learning: anonymized performance data guides default settings and heuristics.

Crucially, stability practices from CAE persisted: adjoint sensitivities remain the backbone for gradient efficiency; MMA variants handle hundreds of thousands of variables; and continuation schedules manage projection sharpness and feature growth. But the target changed from a solitary minimum to a Pareto frontier, and that demands compute orchestration as much as numerical finesse.

Geometry systems enabling the handoff

All the optimization in the world is moot if the shape can’t survive handoff to CAD/CAM. Traditional B-Rep kernels (Parasolid, ACIS) were crafted for precise edges and faces—not for the ragged tessellations and implicit fields common to TO. The response has been threefold. First, mesh/B-Rep hybrids matured: Siemens NX Convergent Modeling lets designers boolean, shell, and dimensionally control facet-heavy bodies alongside precise B-Reps, minimizing translation pain. Second, implicit modeling surged. nTopology, founded by Bradley Rothenberg, popularized field-driven design using signed distance functions and implicit kernels that are resolution-independent and profoundly well-suited to lattices, metamaterials, and post-TO thickening with curvature-continuous blends. These systems trade edge lists for fields and operators, enabling filleting, morphing, and grading as algebra on functions rather than surgery on meshes.

Third, voxel/implicit TO engines evolved with robust meshing and export paths. Frustum Generate, under Jesse Coors-Blankenship and later acquired by PTC in 2018, influenced Creo’s Generative Design with a pipeline that keeps TO tightly coupled to printable geometry and manufacturing constraints. Dassault Systèmes’ CATIA Function-Driven Generative Designer and SIMULIA pipelines connect Abaqus/Tosca with PLM governance on the 3DEXPERIENCE platform, letting field-rich representations coexist with traditional product structures, configurations, and change management. The thematic convergence is clear: instead of forcing optimized fields through brittle reverse-engineering to B-Rep NURBS, tools either embrace hybrid geometry or elevate implicit kernels to first-class citizens, then provide robust tessellation and feature-recognition stages only where needed for downstream CAM, CMM, or drafting.

Beyond classic TO: a broader generative palette

Generative design is not confined to stiffness and mass. Toolchains now weave in multi-physics (thermal, NVH, aerodynamics), reliability margins, and sustainability signals. Heat exchangers exploit field-based porosity control for pressure drop versus heat transfer; NVH targets lead to mode-tuned stiffeners; aero-shaping leverages adjoints to chase drag reduction under manufacturable curvature limits. The optimization core is flanked by surrogate models and design-space exploration platforms: Siemens HEEDS and Altair HyperStudy sample and learn from high-fidelity runs, guiding attention to promising basins of attraction; ML-driven screening filters non-viable regions before expensive solves. In AEC, parametric platforms like Grasshopper (McNeel) and Dynamo (Autodesk) bring goal-seeking and constraint propagation to architectural shells, gridshells, and infill strategies, with TO-informed logic shaping large-scale fabrication strategies.

• Physics breadth: thermoelastic, multiphase cooling, acoustic compliance, fatigue life proxies.
• Sustainability metrics: embodied carbon, buy-to-fly ratios, and repairability indexes.
• Exploration glue: adaptive sampling, Kriging/Gaussian processes, and active learning loops.
• Cross-industry templates: transferable constraint libraries for casting, milling, and AM families.

The net effect is an expanded definition of “generative.” It is no longer merely “topology, then smoothing.” It is a layered decision system where constraints, processes, and economics are peers to structural performance, and where fields and implicit representations provide a lingua franca tying optimization kernels to manufacturing realities.

Conclusion

From mathematical engine to decision workflow

Topology optimization provided the mathematical engine: density or boundary fields governed by elasticity and steered by gradients, filters, and penalization toward material-efficient load paths. Generative design repackaged that engine into a decision-making workflow where users specify intent—loads, materials, manufacturing processes, and economic preferences—and software responds with candidate portfolios instead of a single optimum. The reframe made manufacturability and cost co-equal to stiffness, and elevated the UX from solver parameters to sliders, rankings, and explainable comparisons. Underneath, the same adjoints, MMA moves, and projection operators hum; above, a different orchestration invites collaboration among design, analysis, manufacturing, and procurement. The cultural shift mattered as much as the mathematics: generative tools changed when optimization appears (at concept, not only at verification), who touches it (designers as well as analysts), and how results flow into PLM and ERP systems that govern real products, regulations, and supply chains.

This translation from algorithm to workflow is also a translation from geometry to process. Mesh/B-Rep hybrids and implicit kernels resolve handoff friction; cloud and GPUs tame the need for parallel exploration; and AM-aware constraints convert exotic topologies into practical, certifiable parts. The narrative arc from Michell’s truss lattices to today’s field-driven lattices is not accidental: it reflects a century-long recognition that the right shape is not guessed but computed, and then qualified against the physics and economics of making.

Enablers and competitive dynamics

Three enablers drove convergence on usable, scalable generative design. First, robust optimizers: SIMP with Heaviside projections and length-scale filters; level-set shape derivatives with curvature control; MMA/CONLIN as resilient, massively scalable solvers. Second, new geometry representations: implicit/voxel engines and mesh–B-Rep hybrids that preserve detail and enable physics-aware edits; lattice kernels that treat microstructure as a field, not a tessellation chore. Third, scalable compute: GPUs for interactive steps and cloud orchestration for parallel portfolios. Around these enablers, a competitive landscape crystalized. Altair sharpened OptiStruct and Inspire; Autodesk invested in Fusion 360 Generative Design and lattice technology (augmenting internal research with prior acquisitions); Siemens fused NX/Simcenter with HEEDS; Dassault Systèmes leveraged Abaqus/Tosca within CATIA and 3DEXPERIENCE; PTC folded Frustum’s technology into Creo Generative Design; Ansys blended high-fidelity solvers with interactive Discovery; Hexagon/MSC pushed Apex Generative Design; nTopology led with field/implicit-first workflows. The convergence point was clear: designer-centric UX and PLM integration around rigorous kernels, not toy solvers.

• Algorithms: adjoint gradients, penalization continuation, min–max robustness.
• Geometry: signed distance fields, adaptive tessellation, and hybrid modeling.
• Compute: multi-GPU meshing/solving and elastic cloud scheduling.
• Lifecycle: requirements traceability, variant management, and digital thread continuity.

Next frontiers

Several frontiers are within reach and will determine how deeply generative methods penetrate production. Expect differentiable CAD—kernels that expose stable, analytic derivatives of modeling operations—to replace finite-difference hacks around fillets, drafts, and patterns. Implicit kernels will push toward kernel-level precision, marrying the freedom of fields with toleranced edges and mating features that assemblies require. Standards will matter: richer lattice/field descriptors in 3MF and AMF, STEP AP242 extensions that carry implicit definitions or link to field payloads, and certification-friendly metadata for process parameters and uncertainty budgets. Multi-objective sustainability—embodied carbon, repairability, and supply-chain risk—will join cost and mass as top-tier axes in the Pareto analysis; optimizers will incorporate recycled feedstock variability and energy tariffs as first-class signals.

Equally crucial will be verification and certification loops. As automatically generated designs near mission-critical use, tools must integrate statistically defensible knockdowns, in-situ monitoring correlations, and fatigue scatter models tied to as-built surface states. Surrogate models will evolve from offline approximators to tightly coupled co-pilots that guide sampling, enforce physics-informed priors, and quantify epistemic uncertainty on every candidate. For AEC, the leap will be to construction-aware generative logic that respects modularization, logistics, and site variability. In all domains, the goal is the same: let software explore immensely, but ensure that every selection is traceable, verifiable, and manufacturable—so the promise of generative design translates into everyday, production-grade decisions.