Surrogate and solver, one loop
Structural, thermal and fluid surrogates run the millions of evaluations. Full-fidelity solves are spent where the surrogate is least certain, not on a fixed schedule.
AI Advanced IncProductsNurofusion · PXIstudio
Two products
One searches. One decides.
The model architecture ships as two objects. Nurofusion is the solver — it walks the territory a team cannot reach and returns what survives the physics. PXIstudio is where your engineers stand over the result and choose. We keep them separate on purpose: the moment volume and judgment share a surface, one starts pretending to be the other.
01
The solver
Nurofusion
Learned speed, governed by physics that does not negotiate.
A full simulation is trustworthy and slow; a learned surrogate is fast and, on its own, willing to be confidently wrong about a region it has never seen. Nurofusion fuses the two. Surrogates carry the search at volume, and the governing equations sit underneath as the arbiter — every candidate that matters is checked against the residual before it is allowed to count.
That is the whole of the name. Not a network that has been shown enough parts to guess what a good one looks like, but a learned approximation held on a short leash by the physics it is approximating. The network never gets the last word. When it drifts, the residual says so, and the region is re-sampled at full fidelity rather than believed.
What comes back is not a shape. It is a Pareto set — the candidates that are not beaten on every objective at once — with the trace of how each one was reached, which constraint bound it, and where the search stopped looking.
Residual gating
A candidate passes on the physics residual, not on the model's confidence. Anything the surrogate cannot defend is re-solved properly or dropped.
Twenty objectives at once
Mass, stiffness, thermal margin, build time, cost and qualification burden held simultaneously — against the four or five a person can carry — and returned as a front, not a favourite.
Every run is a trace
Seeds, constraint set, solver versions and the whole descent are kept. A result you cannot reproduce in front of a reviewer is not a result.
Nurofusion · at a glance
| Dimension | What ships |
|---|---|
| Physics | Structural, thermal and fluid, steady and transient. Coupled where the problem is coupled. |
| Search | Topology optimisation, generative geometry and compositional search, with manufacturability as a constraint inside the loop rather than a downstream surprise. |
| Inputs | The brief as a formal object — objectives, hard constraints, soft preferences, success criteria — plus your existing CAD, load cases and material data. |
| Outputs | A Pareto set with geometry, evaluated performance, the binding constraint per candidate, and the full optimisation trace. |
| Runs | On your infrastructure or ours. Nothing about a programme has to leave your environment for the search to happen. |
02
The studio
PXIstudio
Where ten thousand viable answers become one decision, with a reason attached.
A search that returns ten thousand valid candidates has not finished the work — it has moved the work. Somebody still has to say which one is right for these people, in this place, at this moment, and then defend that in a review. PXIstudio is the room where that happens.
It opens on the front rather than on a winner. You can pull the trade in any direction and watch what it costs elsewhere, ask a candidate why it survived, and see the constraint that bound it. Unfamiliar geometry is the normal case here, so intuition is not the sanity check — the evidence is.
The studio never chooses. It makes choosing legible: what each candidate optimises, what it sacrifices, where the model may have been blind, and what your team noticed that the model could not. Then it writes that down as an engineering artefact, signed by the person who decided.
01 OpenStart from the front, not a winner
The Pareto set loads whole. Nothing has been pre-selected on your behalf, because a hidden default is an unstated objective wearing a helpful face.
02 CompareMove the trade and watch what it costs
Weight one objective harder and the front redraws. The point is not to find the answer but to see the shape of the compromise before committing to one corner of it.
03 InterrogateAsk a candidate why it survived
Load paths, margins, the binding constraint, the process window it assumes, and the fidelity of the evidence behind each number — full solve or surrogate, stated plainly.
04 AnnotateAdd what the model could not see
The technician who has to reach it. The supplier you cannot change. The reason the last design failed review. Tacit constraints get written onto the candidate, and can be pushed back into the brief for the next run. Human-led.
05 DecideChoose, and own it
One candidate is selected by a person whose name goes on it, with the runners-up and the reason for setting them aside preserved beside it. Human-led.
06 ExportHand over something a reviewer accepts
Geometry, the evidence package, the trace and the decision record leave together — so an unfamiliar shape arrives at design review with its argument already attached.
03
Why they are two things and not one.
The split is the argument the company is making, drawn as software.
Volume moves to the model. Judgment does not.
Nurofusion is measured on how much of the territory it reaches. PXIstudio is measured on how clearly a person can decide. Those are different jobs, and a single tool that claimed both would quietly do the second one for you.
One artefact runs through both
The trace the solver writes is the same trace the studio reads, and the decision the studio records is the same object the next brief starts from. There is no export step where the reasoning falls out.
Either one, on its own
Teams with their own solver stack use PXIstudio to make a front reviewable. Teams with a review process they trust take Nurofusion alone. Most start with one problem and one product, which is the honest way to find out whether this works for you.
“When a model proposes ten thousand viable answers, taste matters more than productivity.”