Turn the design into variables
Parametric geometry with every variable declared: unit, bounds, default and a clear refusal when a value makes no sense. The design space becomes something you can explore instead of redraw. Design exploration
Home / Product / TensorLynx
Product · TensorLynxTensorLynx: simulation, optimisation and AI that prove their answers
Fast predictions are easy to buy. Answers an engineering lead can sign are not. TensorLynx is OLTRION's engineering platform that takes a design from parametric geometry through coupled fluid, structural and thermal simulation, adjoint and multidisciplinary optimisation and trust-aware AI to a design that has been re-checked on the full model, with every step on record.
Bars set before measurementUnsure predictions withheldVerified only after an independent re-run
The trust gap
The industry has adopted AI for simulation. It has not learned to trust it.
Every vendor now sells speed. Engineering teams tell surveys a different story: the hard part is knowing when a fast answer is right, and turning a promising pilot into something a programme can rely on.
85%
of organisations surveyed have begun testing or deploying AI in their simulation workflows
26.2%
report operational satisfaction with it. Trust and validation of results is named as a barrier by 47%
16%
of manufacturers surveyed trust AI-supported verification and validation
92% vs 8%
of engineering leaders deployed surrogate models in 2025, yet only 8% say they are extensively deployed in their workflows
The gap is not computing power. It is evidence: a prediction that cannot say how sure it is, a gradient nobody checked, an optimised shape that was never re-run, a result no one can reproduce six months later. TensorLynx was built to close exactly those gaps.
Our thesis
Speed is solved. Trust is engineered.
We did not start from a model and look for problems. We started from the questions a design review asks, and built the platform so every answer arrives with its evidence. Six rules run through every workspace, every study and every report.
They are not marketing lines. They are how the software behaves: the check limits are written into the platform before a single run, and the platform refuses to round an unknown into a yes.
- 01
Bars before measurement
Every pass or fail limit is declared with its reference before a result exists, and is never moved after one does.
- 02
Unknown is never a pass
A check that was not measured reads Unknown. Four passes and one unmeasured check is not a pass.
- 03
Missing is never zero
A failed run is recorded as failed, with its outputs missing, never as a plausible number that quietly enters a model.
- 04
Withhold rather than guess
When a surrogate fails its confidence checks, its prediction is hidden and the decision falls back to solver results.
- 05
Verified means re-run
A design is called Verified only after an independent re-run on the full model, not because an optimiser liked it.
- 06
The engineer decides
The platform ranks, explains and recommends the best evaluated feasible design. The decision stays a recorded human act.
How it works
One loop from the first sketch to a design you can defend
Most toolchains hand results from one program to the next and lose the context on the way. In TensorLynx each stage passes its result to the next on the same model, in the same units, into the same record.
Solve the physics together
Flow, structure and heat exchange data at their interfaces, with conservation checks and pre-flight refusals that name the fix. Coupled multiphysics · Mesh verification
Ask the design where to move
Adjoint solves give the sensitivity of the objective to the whole shape at once, and a gradient route must pass a finite-difference check before it may steer a new case. Adjoint optimisation · Gradient verification
Learn from every solve, honestly
Surrogates and full-field reduced-order models trained on your own simulations, each reporting its confidence and stepping aside when it is not good enough. Trust-aware surrogates · Full-field ROM · Multi-fidelity
Respect every constraint at once
Multidisciplinary studies built from declarations: objectives, constraints and disciplines with their maturity, so feasibility is designed in rather than checked at the end. MDAO
See the trade-offs, choose deliberately
Fronts, knees and rankings make the compromise between objectives visible, and the choice is recorded as a decision with its reasons. Pareto selection
Re-run before recommending
The chosen design is re-solved on the full model. Only then is it called Verified, and the gap between prediction and full model goes into the report. High-fidelity reassessment
Keep the evidence
Every run seals what produced it, and a report will not export a claim that cannot be traced. The next cycle starts from the record, not from memory. Provenance
Twelve capabilities, one platform
Every chapter of the loop, in depth
Each capability has its own page: the engineering problem it solves, how TensorLynx does it, the evidence behind it and how it hands over to the next step.
Geometry-driven design exploration
Parametric geometry that turns every design question into a study, not a redraw.
Coupled fluid, structural and thermal engineering
Fluid, structure and heat solved as one exchange, not three separate reports.
Mesh refinement and numerical verification
Numbers you can defend: refinement studies and discretisation error, on record.
Adjoint-based shape optimisation
One adjoint solve, the sensitivity of the objective to every surface point.
Sensitivity analysis and gradient verification
Gradients checked before they steer a design.
Surrogate modelling and uncertainty assessment
Fast predictions that say how sure they are, and stay silent when they are not.
Full-field reduced-order prediction
Whole pressure, velocity and stress fields from the designs you have already solved, with their error shown.
Multi-fidelity optimisation
Spend expensive solves only where cheap ones cannot decide.
Multidisciplinary design optimisation (MDAO)
Objectives, constraints and disciplines declared once, and a recommendation that respects them.
Pareto-based candidate selection
Trade-offs made visible, so the choice is an engineering decision, not a guess.
High-fidelity reassessment of optimised designs
No optimised design is recommended until the full model has confirmed it.
Reproducibility, provenance and engineering history
Every result traceable to the geometry, mesh, solver and settings that produced it.
Inside the platform
A vocabulary that cannot overstate a result
TensorLynx labels every check, every value and every prediction with a word that means one thing. A reviewer reading a TensorLynx report knows what was measured, what was not, and what was deliberately withheld.
- Verdicts come from declared rules with references, never from judgement after the fact.
- Value status keeps failed, missing and unverified data out of models.
- Prediction labels tell you whether a surrogate is inside the data it was trained on, at its edge, or outside it.
The workbench
Six workspaces. One model. One record.
TensorLynx runs in the browser. Geometry, simulation, optimisation, AI, results and history are workspaces of one application, not separate programs joined by file exports.
- Geometry reads and writes standard CAD exchange formats (STEP, IGES, BREP) and keeps meshing isolated from the application.
- Every parameter is declared once, with unit, bounds and default, and the interface builds its controls from that declaration.
- One study model drives different solvers from declarations; disciplines come from a registry that states their maturity, and roles are assigned, never guessed from variable names.
- Reports are self-contained files with a "Verification evidence and limitations" section and a recorded decision.
Evidence you can check
Recorded runs, with their caveats on the same page
These are measured results from recorded TensorLynx runs on internal engineering cases. Each states what it was measured on, because a number without its context is exactly what TensorLynx was built to avoid.
Load-bearing bracket
61.1% lower
Aggregated von Mises stress measure (KS) from 24.4 to 9.50 MPa at constant volume; 59.0% lower when re-solved on a finer mesh, with a Grid Convergence Index of 0.13%.
An aggregate stress measure, not a literal peak-stress value.
S-duct intake
8.0% lower loss
Pressure loss from 79.30 to 72.98 Pa across a CAD-parametric family; 8.9% on a second mesh. The design exported back to CAD and read back to a volume difference of 1.46e-14.
One operating point; the best design sat at the edge of the parameter box in three of four parameters.
U-bend, coupled heat transfer
0.0023% closure
Fluid and solid solved as a partitioned coupled problem: the heat balance closed to 0.0023% in 7 coupling iterations.
Coupled analysis; coupled optimisation was not part of this case.
A surrogate that stepped aside
Withheld
On an airfoil study the surrogate's uncertainty covered only 21% of held-out results against a 95% target. TensorLynx hid its prediction, and the recommendation came from solver results alone.
This is the behaviour we want buyers to see: the platform refused to use a model that had not earned it.
Where it applies
Built for design decisions that must be defended
Aerospace and defence
Intakes, ducts and aerodynamic shapes refined by adjoint sensitivities; structural sizing where the verification record matters as much as the result.
Automotive and e-mobility
Cooling paths, brackets and thermal-structural trade-offs explored across families of designs instead of one candidate at a time.
Energy and turbomachinery
Pressure losses, heat exchangers and coupled thermal loads, where a few percent of efficiency is worth a careful, checked optimisation.
Electronics cooling
Heat sinks and cold plates evaluated with conjugate heat transfer over the operating points that matter, weighted the way you run them.
Process and industrial equipment
Manifolds, fittings and duct systems where surrogates speed up studies and every recommendation is re-checked on the full model.
What changes
From a chain of tools to one accountable loop
| Question in the design review | A typical toolchain | With TensorLynx |
|---|---|---|
| Is this AI prediction good enough? | A number, often without a confidence measure | A label and confidence, or a withheld prediction and a solver run |
| Can we trust the gradient that drove the shape? | Assumed correct | Checked against finite differences; "Direction-checked" is never sold as "Verified" |
| Was the optimised design re-checked? | Sometimes, by hand, later | Re-run on the full model before it is called Verified |
| Is the mesh good enough? | Judged by eye | Refinement studies with a Grid Convergence Index on record |
| What happened to the failed runs? | Dropped, or filled with a default | Recorded as failed; never turned into zeros in a model |
| Can we reproduce this in a year? | If someone kept the files | Sealed run records; reports refuse untraceable claims |
Questions engineers ask
Before you talk to us
If your question is not here, ask an engineer directly. You will talk to someone who runs TensorLynx on real work.
What is TensorLynx?
TensorLynx is OLTRION's engineering-simulation platform. It brings parametric geometry, coupled fluid, structural and thermal simulation, adjoint and multidisciplinary optimisation, AI surrogates and verification into one web workbench, so a design moves from question to verified answer on one model and one record.
How is TensorLynx different from adding AI to an existing simulation toolchain?
The AI sits inside a verification loop rather than beside it. Pass and fail bars are declared before anything is measured, an unmeasured check is reported as Unknown and never as a pass, a surrogate prediction that fails its confidence checks is withheld instead of shown, and a design is only called Verified after an independent re-run on the full model.
Does TensorLynx replace physical testing or certification?
No. Simulation and AI predictions support engineering decisions; they do not replace the tests, approvals and sign-off a product requires. What TensorLynx adds is a traceable record of how each result was produced, which makes the engineering case easier to review.
Can we start without buying software?
Yes. Most work starts as a feasibility review or a study that OLTRION engineers run on your geometry with TensorLynx. You receive the verified results and the full record, and you can move to a workflow built for your product or to platform access for your team later.
What happens to our geometry and data?
Before any work starts we agree with you how your geometry, results and models are stored, shared and returned. Each TensorLynx run records what it used, so you can see exactly which inputs produced which result.
Which industries is TensorLynx built for?
Engineering teams who design flow paths, structures and thermal systems: aerospace, automotive and e-mobility, energy and turbomachinery, process and industrial equipment, and electronics cooling. The common thread is a design decision that must be defended with evidence.
One loop, not twelve tools
The TensorLynx engineering loop
Every capability hands its result to the next on the same model, the same units and the same record: geometry feeds physics, physics feeds sensitivities and learning, optimisation proposes, verification confirms, and the history keeps every step reproducible.
Work with OLTRION
Bring us the design question. Leave with a verified answer and the record behind it.
TensorLynx is built and run by the engineers who use it on client work. You can start small, with one geometry and one question, and grow into a workflow your own team owns.
Feasibility review
Send a geometry, the objective and the constraints. We tell you which TensorLynx route answers it, what it needs from you, and what you will get back.
We run the study
Our engineers run the loop on your problem: physics, sensitivities, optimisation and full-model verification, handed over with every setting on record.
A workflow built for your product
A TensorLynx workflow, surrogate and verification gates set up around your product family, so the next design cycle starts where the last one ended.
Platform access and handover
Your team on TensorLynx, trained on your own cases, with OLTRION engineers behind it for the questions that need a specialist.
Simulation and AI predictions support engineering decisions; they do not replace the testing, approval and sign-off a product requires. Figures marked measured come from recorded TensorLynx runs on the stated case; figures marked synthetic or schematic illustrate a method.