Guide

Generate design variants that respect your constraints

Exploring a design space means building each variant, and that is the step that caps how many you ever try. A generative model learns the geometry from designs you have already made, including the features that must not move, and produces new variants in seconds that keep those constraints intact and stay ready to evaluate.

Read this page top to bottom and it runs the pipeline in order: the designs you already have go in, the constraints inside them get learned, and new variants come out the other end as candidates for your normal evaluation.

Request a Demo Generative Design

The bottleneck

Why variant count stays low

Each new geometry is manual CAD work, so the design space you actually evaluate is the small corner you had time to build.

The limit is rarely the solver and it is never the idea list. It is the hours between a sketch and a body clean enough to mesh. Teams answer that by choosing a handful of variants early and defending them, which quietly turns an exploration into a confirmation of the first guess. Everything outside that corner stays an opinion, because nothing ever attached a number to it.

Input bracket design one, a 3D-rendered mounting bracket with boreholes. — Step 01, what goes in. One of the bracket designs the model is trained on, a part that already exists.

01 Input designs

A shape family, not a parameter list

Trained on a handful of geometries, it learns the shape family rather than a parameter list, which is why the variants it produces look like your parts and not like noise.

A parameter list can only move the dimensions somebody thought to expose, and it stops at the edge of that list. A learned shape family carries the proportions, the transitions, and the features that repeat across parts you have already signed off. That is the difference between generating a new bracket and generating a shape that happens to have a hole in it.

02 Learned constraints

Constraints are learned, not bolted on

In a generative geometry example, boreholes were learned as constraints from three input designs, with training in under a minute on a GPU and new variations produced in seconds. One example, one setup: a different part family will answer differently.

Nobody told the model those holes were bolt positions. They held because they held in every design it was shown, so it treated them as part of the family rather than as something to vary. That gives you a practical test for what will survive: a feature that is genuinely invariant across your training geometries comes through the generated ones, and a feature that drifts between them will drift in the output too.

Three

Input designs the example learned from.

Under a minute

Training time on a GPU in that example.

Seconds

Time to produce new variations once trained.

03 Generated variants

What comes out the other end

Many shapes from the same family, produced in seconds rather than sketched one at a time. They are candidates to evaluate, which is exactly what a wide search needs at this stage.

A grid of generated bracket geometry variants grouped into color-coded clusters, produced by STOCHOS from a few input designs. — Step 03, the payoff. Generated variants grouped into clusters, so you can pick across the space instead of picking near-duplicates.

In your process

A wider shortlist, the same sign-off path

Generated variants feed the same evaluation you already run, and every geometry that matters still goes through your normal CAD and validation path before anyone makes it.

Treat the output as a shortlist of candidates, not as finished parts. Screen them the way you screen anything else, with a solve or with a surrogate model, and take the few that survive into CAD. The gain is the width of the search that happens before that point, not a step removed from the end of it. The method behind this is set out on Generative Design, and STOCHOS is the engine that runs it on your own infrastructure.

Common questions

01Are the generated geometries ready to manufacture?

No. They are candidates to evaluate. Your normal CAD, validation, and sign-off path still decides what gets made.

02How many existing designs does generative design need?

Fewer than most people expect. The generative geometry example learned from three input designs. How far that generalizes depends on how varied your family is.

03Can generative design hold hard constraints like bolt positions?

That is the point of learning constraints from the input designs. Features that are fixed across the training geometries are carried into the generated ones.

Method

Generative Design

The method page: learning a shape family, generating geometry, ranking what comes back.

Method

Surrogate Modeling

How to screen a long candidate list without paying for a solve on each one.

Product

STOCHOS

The predictive engine behind the generated geometry, running on your infrastructure.

Index

News and Guides

The rest of the guides, plus what we have published recently.

Next step

Show us a part family

A few geometries you have already built are enough to start the conversation. We can look at which features are invariant across them, which ones are not, and what a generated set would give you to evaluate.

Request a Demo Generative Design

Local-first. Your data stays on your infrastructure.