Generative design shapes
A Socratic walk-through of generative design shapes — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #Why do computer-evolved brackets end up looking like bones rather than beams?
Hand a topology optimiser a mounting bracket problem — these bolt holes are fixed, this load arrives here, use no more than a fifth of the available material — and it returns something that looks unsettlingly organic. Swept curves, branching struts, hollow cavities, thickened junctions. It looks like a shoulder blade or a section of trabecular bone. It does not look like anything a draughtsman would draw.
The romantic reading is that the software has discovered nature's designs. The dismissive reading is that it is a rendering artefact. Both skip the interesting question, which is why an algorithm that knows nothing about biology, given only a load case and a material budget, should converge on the same visual vocabulary that vertebrate skeletons use.
Reasoning it through
REASONING #Start with what the optimiser actually does, because the word "evolved" in the question is doing some misleading work. Most commercial generative design is not a genetic algorithm breeding populations of brackets. The standard method fills the whole allowed volume with a notional material whose density can vary continuously between empty and solid, solves for how stress flows through it, and then nudges each element's density up or down according to how much that element contributes to overall stiffness. Repeat a few hundred times, with a penalty pushing intermediate densities toward fully solid or fully absent, and the field resolves into a shape.
But notice the shape of that loop, whatever we call it. Each iteration asks of every scrap of material: are you carrying load? Material that is carrying load is retained and reinforced. Material that is barely stressed is thinned toward nothing. That is a selection process. It does not derive the shape from first principles or search a catalogue of forms — it culls, repeatedly, against a single criterion, and the shape is the residue of what survived. Some variants make this literal: the evolutionary structural optimisation family, developed in the early 1990s, simply deletes the least-stressed elements each round.
Now ask what a selection process like that must produce. Where does load actually go in a solid? It flows along paths from where it is applied to where it is reacted, and those paths are curved, because the shortest stiff route between two anchor points under a bending load is not a straight line through material that is doing nothing. Material off those paths contributes stiffness far below its weight cost, so it is removed. Material at a junction where several paths converge must carry all of them, so it thickens. Material in a region loaded from more than one direction stays as a shell rather than a solid. Run this to convergence and you get curved struts, fillets at junctions, hollow sections, and lattice in between.
So now the biology. Bone is subject to a strikingly similar loop. Bone tissue is continually remodelled: osteoclasts resorb it, osteoblasts deposit it, and the balance between them responds to local mechanical strain. Regions that are strained get reinforced; regions that are unloaded are resorbed — which is why bed rest and spaceflight cause measurable bone loss. The idea that bone adapts its architecture to its loading goes back to Wolff in the 1890s, and the modern quantitative version, the mechanostat model, is due to Frost.
Put the two side by side. Same physics of load flow through a solid. Same objective, roughly: stiffness per unit mass, because mass is expensive both to a satellite and to an animal that has to carry it. Same mechanism: retain what is stressed, remove what is not, iterate. Convergent form is then not a coincidence and not mysticism — it is what you should expect when two different processes run the same selection rule against the same constraints.
Two honest cautions. The resemblance is real but it is not identity: bone is also a mineral reservoir, a marrow housing, and self-repairing, and it grows under developmental control that has nothing to do with stress. And the optimiser's output is smoother than the physics requires — much of the "organic" look comes from mesh smoothing and from running the problem with no manufacturing constraints at all. Add a requirement that it be millable in three axes, or castable with a draft angle, and the same solver returns something considerably more like a beam with holes in it.
The analogy
THE ANALOGY #Think of a desire path across a lawn. Nobody designs it. Each walker takes the route that costs them least, the grass where feet fall dies back, the bare strip is then slightly easier to walk on, and the path emerges as the residue of thousands of individual selections. It curves where the direct line would be steep or muddy, it widens at junctions, and it looks nothing like the paved right angles the landscape architect drew.
desire paths are shaped by many independent walkers each optimising their own crossing, whereas a topology optimiser evaluates the whole structure against one global objective every iteration — so a desire path can end up locally sensible and globally poor, while the optimiser cannot, and that difference is exactly why the algorithm converges and a lawn does not.
Clarifying the model
THE MODEL #Three refinements.
First, the misconception the question invites: the software is not imitating nature. It has no biological reference, no library of skeletons. The resemblance runs the other way — both arrive at similar forms because both are selection processes over the same load physics, and neither is copying the other.
Second, "computer-evolved" should be held loosely. Gradient-based density methods and genuinely evolutionary methods produce comparable-looking results, which is itself informative: the organic appearance comes from the selection criterion and the physics, not from the particular search technique.
Third, the shapes are honest but not free. Unconstrained topology optimisation routinely returns geometry that cannot be machined and is awkward even to print, and it is optimal only for the load cases you specified — feed it one load and it may be startlingly fragile under a second one you forgot to mention. Bone, which is loaded from every direction its owner moves in, has a much richer specification than most brackets are given.
A picture of it
THE PICTURE #How to readStart at the input at the top and follow down to the diamond — that single question, asked of every scrap of material, is the entire selection rule. Both of its branches feed the junction, and the labelled edge from the junction back up to the stress solve is the loop that does the real work; the bone-like form is what remains after a few hundred passes around it. The final node is the standing hazard: the residue is optimal only against the load cases and constraints that were actually stated.
What became clearer
WHAT CLEARED #The bone-like look is not an aesthetic and not an imitation. It is the signature of a selection loop — keep what carries load, discard what does not, repeat — run against the way stress actually flows through a solid. Bone remodelling runs the same loop with cells instead of elements, which is why two processes with nothing in common except their physics and their criterion arrive at the same vocabulary of curves and branches.
Where to go next
ONWARD #- Michell's 1904 optimal truss layouts, which reached the same forms analytically long before computers.
- How manufacturing constraints reshape optimiser output, and why additive manufacturing revived interest in the field.
- Multi-load-case and robust optimisation, and what happens to these shapes when uncertainty in the loading is priced in.
Key terms
TERMS #| Term | What it means |
|---|---|
| Topology optimisation | computing where material should and should not be within a given volume, subject to loads and a material budget. |
| Compliance | the inverse of stiffness; minimising it is the usual objective here. |
| Density penalisation | the numerical trick pushing intermediate densities toward solid or void, so the answer is a shape rather than a grey fog. |
| Bone remodelling | the continuous resorption and deposition of bone tissue, responsive to local strain. |
| Mechanostat | Frost's model of the strain thresholds at which bone is added or resorbed. |
Every term the collection defines is gathered in the glossary.