My first topology optimization run was a triumph on screen and a disaster on the shop floor. I had a 4.2 kg bracket, and the solver happily carved it down to 2.3 kg with a shape that looked like the skeleton of a sea creature. The sales team loved the pictures. Then I sent the result to the machining shop and the quote came back at four times the price of the original part, because that beautiful topology had to be held in five setups with three custom fixtures, and half its webs were too thin for any standard end mill. I kept the screen capture; I also kept the lesson. Since then I have run a lot of lightweighting projects, and the ones that survived contact with reality all followed the same bones. This is that workflow, warts and all.
Step one: the load case is the whole game
I cannot stress this enough. Nine out of ten wasted optimization projects fail before the solver runs a single iteration, and the reason is a load case that does not describe the real machine. If your part actually sees a 3 g shock when the index table stops, and you only simulate a 1 g static gravity load, the optimizer will remove exactly the material that keeps you alive during that shock. A few practical numbers from tooling components: an indexing table stopping a 40 kg turret from 60 rpm in 0.15 s generates a tangential force you should not ignore, and a robot end effector spinning at 2,800 rpm has centripetal loads on the fingers that dominate everything else.
So before anything else I write down the load envelope with the shop stuck on my shoulder: maximum static force, the dynamic corner, the temperature case if the part gets hot, and — the one everyone forgets — the assembly and maintenance loads. That last category has killed more optimized parts than fatigue ever did. A part can be perfectly sized for service and snap the first time a fitter stands on it to reach a bolt.
Optimization does not create strength; it distributes the strength you have. If your load case is wrong, the extra strength gets distributed to the wrong places.
Step two: symmetric models and honest constraints
Topology optimization rewards symmetry because it halves the solve time and often a quarter. But there is a trap: forcing symmetry in a model that is not symmetric in reality gives you a part that is symmetric and wrong. My rule is to apply a symmetry plane only when the geometry, the loads and the constraints are all genuinely symmetric — and to keep the factory constraint (how the part is bolted to the machine) exactly as it will be. Ship mounting points as non-design space, always. A topology result that removes material around a bolt hole because the solver decided the flange was optional has misread your intention: you need that flange for assembly, service and energy transfer, not just for structural load.
I also keep the manufacturing constraints in the setup from the very first iteration: minimum web thickness, draft direction for casting or the tool access direction for machining, and the material I intend to use. A part optimized as if it were machined from billet looks very different from a part optimized for a casting with a draft of 1.5 degrees. Getting that right early is the difference between a topology you admire and a topology you manufacture.
Step three: reading the density plot like a normal person
The colour map is not a wall decoration. In a classic density-based solver, red means “definitely keep material” and blue means “the solver would happily throw this away”. The trick is knowing where to cut the threshold. I usually look at the 0.3 contour first — everything below that density is noise I will never machine, and everything above it is the load path. The single best habit I have is to ask, before touching CAD: “Does this result match my engineering intuition?” If the topology is carrying load through a path I would never have guessed, I investigate, because either the load case is lying to me or my intuition was. Both are worth knowing.
An honest warning about the checkerboard problem. Old solvers and badly set meshes spew out striped, salt-and-pepper density fields that look scientific and are actually numerical noise. If your result looks like a fingerprint or a badly tuned radio, refine the mesh and check the mesh quality before you believe a single red cell. I have shown enough clean, believable topologies to know the real ones look deliberate.
Step four: turning topology into machinable geometry
This is where the sea-creature problem lives, and the fix is brutal simplification. I never machine the raw topology. I machine an interpretation of it that a tool can actually reach. The comparison below has saved me countless meetings:
| Design option | Machinable? | Weight | Real cost in my shop |
|---|---|---|---|
| Raw topology mesh | No — 5 setups, custom fixtures | 2.3 kg (ideal) | 3 to 4X a regular part |
| Interpreted with straight webs and big fillets | Yes — 2 setups, standard end mills | 2.9 kg | 1.1 to 1.3X a regular part |
| Original solid bracket | Yes — 1 setup | 4.2 kg | baseline |
You will notice I did not reach the theoretical minimum weight. I do not care. The interpreted part keeps about 70 percent of the weight saving and is a boring, routine machining job. That trade is the entire profession: the winning geometry is the one the shop quotes without a sigh, not the one that wins the simulation screenshot.
When I rebuild the geometry I keep four rules visible on my monitor: make webs no thinner than 6 mm for tall pockets, keep every pocket reachable by a standard 20 mm end mill, give every inside corner at least a 3 mm radius or plan the fillet as a feature, and leave the mounting flanges exactly as deep as the bolts and the service access need. The result is a part an apprentice can machine on the second try, which is the humility the sales team never asks for.
Step five: size optimization after topology, not instead of it
People treat topology and size optimization as rivals. They are not; they are two passes of the same job. Topology decides where material should exist; size optimization decides how thick each web and wall has to be. I always follow a topology pass with a size pass on the interpreted geometry, sweeping the web thickness as the design variable and watching the stress and the mass trade off against each other. On one clamp housing this second pass took the part from 3.7 kg down to 3.2 kg without touching the topology, just by thinning webs that had honestly been guesses when I rebuilt them.
The sequence looks like this, and it fits on one slide:
- Solve the topology with realistic load cases and manufacturing constraints.
- Rebuild in CAD as clean, machinable features.
- Run a size optimization on wall and web thickness over a realistic range.
- Verify the final part statically, then check the dynamic corner.
- If the part sees repeated loading, run a fatigue check with the real load spectrum before release.
I have seen teams skip step five and release a lightweighted part that shakes itself to pieces at resonance in a month. The lighter you make a part, the more its natural frequency drops, and the more interesting it becomes to the vibration environment it lives in. A bracket that was plain solid at a 140 Hz natural frequency can slide down to 95 Hz after aggressive lightweighting — and if the machine has a pump at 100 Hz, you have built a tuning fork.
Pitfalls that cost real money
| Pitfall | Symptom | My fix |
|---|---|---|
| Simulating the prettiest load case, not the daily one | Part survives the report, fails in service | Load envelope written by the shop and the maintenance crew |
| Forgetting assembly and maintenance loads | Fitter breaks the optimized part reaching for a bolt | Add a 1.5 safety note and simulate the service position |
| Believing the raw topology | Unmachinable geometry, scary quotes | Always interpret to features a standard cutter reaches |
| No natural frequency check | Resonance, fatigue, mysterious failures at 3 a.m. | Modal check before release, re-size if it drops near running speeds |
| Zero fatigue pass on a cyclically loaded part | Cracks along the new thin webs | Fatigue analysis with the real load spectrum |
| Material changed after optimization | Aluminium part optimized for steel stiffness goes floppy | Lock the material before the first solver run |
Closing thoughts from a guy who married the sea creature
Lightweighting is a team sport that includes the shop, the maintenance crew and the procurement person who has to source the material. The practical payoff is real — I have seen a transfer line save tens of kilograms on every pallet by optimizing the pallet structure, and the acceleration time of the servo dropped noticeably because of it. But every win came from a workflow that treated manufacturing as a constraint from iteration one, not as a complaint at the end.
Optimize to the shop, not to the screenshot. The screen will always show a prettier shape; the shop will always build the honest one. Find the geometry in between, and you have done the actual job. That is the part no animation ever shows, and it is the part that pays the bills.
A lightweighting project, from quote to floor
A customer came to us with a rotary indexing platform that was too heavy for the servo they had to use. The platform was 620 mm across, cast, and weighed 48 kg, and the servo they wanted needed it under 40 kg to hit their cycle time. The obvious move was to mill big lightening pockets and call it done, but that is how you get a platform that flexes 0.3 mm under tool load and ruins every workpiece it carries. So instead we set up the model with the real duty: an 1,800 rpm spindle motor bolted on, a 3.2 kg pallet with fixture, a worst-case tool engagement force of 1,400 N at 270 mm radius, and the clamping points exactly as the machine mounts them.
The topology pass carved the bottom face into a web pattern that no one on my team predicted, and honestly the first interpretation was ugly. But the second interpretation, with 8 mm webs and a full-depth central ring to keep the stiffness where the spindle mounted, came out at 41 kg and passed the deflection check at the worst-case position with 0.09 mm total deflection against a 0.15 mm limit. The vendor quoted normal numbers because every pocket was reachable by a 20 mm end mill from one side. The servo saw the mass it wanted.
What surprised me most was the measured result: the actual deflection on the shop floor was very close to the simulation, which told me the load case had been honest. That is the quiet recipe, and it never gets a press release — build the load case like you are the part, optimize to the tool, and validate with the same physics that made the numbers in the first place.
When not to lightweight at all
I should also say the obvious thing nobody likes to hear: not every part should be optimized, and the machine tool world is full of parts that are intentionally heavy because mass is a feature. A mill column that is stiff because it is heavy beats a light column that dances. A machine base that absorbs vibration by being massive beats a base that transmits it. Lightweighting wins where the mass moves, where it has to be accelerated, or where it is bolted to something that already flexes. If the part is structural, static and happy being heavy, optimizing it is still great fun and still a mistake. I have stopped more optimization projects than I have started, and I am proud of that number.
The workflow in this article is for the parts that genuinely travel, swing, or get accelerated. If you are sure your part is one of those, set the loads honestly, constrain the manufacturing, interpret the result with a standard cutter in mind, and check the dynamics before you release. Do that, and the optimization will earn its keep in cycle time and energy, not just in a slide deck.
