No pages match that search.
Product / TensorLynx / Design exploration
Define · Geometry-driven design explorationMost design questions still begin as a redraw: edit the CAD, re-export, re-mesh, and hope the boundary names still point at the right faces. The TensorLynx platform turns the question into a study. You declare which dimensions may move and how far; every variant is regenerated from CAD, checked, meshed on its own and solved, and nothing that fails is quietly lost.
Design variables carry type, unit and boundsNamed faces resolved by rule, never by indexSeeded, reproducible sampling plans
Schematic Drawn from the S-duct family's station formula; move the pointer over it.
Hand-built variants cost analyst hours, and everyone budgets for that. Nobody budgets for a variant that runs cleanly and is wrong. We measured three such failures while building TensorLynx; each is now stopped before it reaches a solver.
The mesh still builds. The solver still runs. The answer belongs to another design.
A reference script named its boundaries by the CAD kernel's face numbers. Meshed at five duct diameters, the numbers survived and their meaning did not: at three of the five, the wall group picked up a face from the half the symmetry cut had removed. Measured · in-house, 5 diameters
On an imported B-rep box, every requested maximum size from 20 down to 5 gave a byte-identical mesh of 988 elements. A refinement study built on it still produces a number. Measured · 30×20×15 box
A Sobol' sequence keeps its balance only at a power-of-two count. Ask for 12 points instead of 16 and part of the design space is never visited, with no error raised. Method property
None of these appear in a residual plot. They surface later, as a decision made on a design nobody analysed.
Give a dimension a minimum, a maximum and a unit and it becomes a design variable, the only kind a study may move. An inverted range, or a nominal value outside its own bounds, is refused when the variable is created; a zero-width range is refused with the fix named. Inside, everything is SI.
Latin hypercube, Sobol' or full factorial, drawn in the normalised unit box. The seed is fixed by default, because a plan whose points change between runs is not a study. Ask for a Sobol' count that is not a power of two and the platform names both neighbours rather than rounding silently.
Each sample point is rebuilt by the B-rep CAD kernel in memory. The result must prove it is still a member of the family: every named face resolves to the number of faces it declared, and every face is claimed.
Every design gets its own mesh in an isolated meshing worker, sized from stated targets, a y⁺ and cells across a reference length, not mesher settings. Named faces become mesh markers by rule, and each is cross-checked by face area before anything is named.
A solve becomes data only when its objective has stopped moving. A design that fails to build, mesh or solve is recorded with stage and reason; one that will not settle is left out, reason shown. Missing is never zero, and a resumed campaign re-solves nothing.
The campaign opens in Results Analysis and History like any solver run. What it often shows first is the box itself: where the response is flat, where it is steep, and whether the best region is pressed against a bound you set. That is the moment to widen a range.
Schematic S-duct family bounds as declared; sample positions are an unscrambled Sobol' projection; response colours are a Synthetic demo. The bound in step 6 is the measured result shown under Evidence.
Move the family's box yourself. The sliders carry the S-duct family's declared bounds, and the drawing follows its station formula: the centre line rises by a smooth step of u = fskew, the mid-section diameter scales by throat along a 4f(1 − f) bump, and the section keeps its area while its aspect changes with flat.
DesignVariable(name, value, minimum, maximum, units) · xi = (vi − mini) / (maxi − mini) ∈ [0, 1]
Bounds live beside the value, so stored documents keep loading and "no bounds" stays a normal state for a parameter that should not move. The readiness check lists, in one pass, every reason a document is not yet a family, such as a design variable without bounds or a face no region claims.
Sobol': balanced at n = 2m · full factorial: levels = n1/d, rounded and reported
Latin hypercube sampling stratifies each variable into n intervals sampled once; Sobol' points are a low-discrepancy sequence balanced at powers of two. A shape-family campaign plans the nominal design, then a scrambled Sobol' sequence over design variables and operating conditions together, a pure function of the specification, so a resumed run asks the same questions.
| Verdict | Meaning | Effect |
|---|---|---|
| VALID | the same faces as the last resolution | proceeds |
| REMAPPED | the declared number of faces, but different ones | proceeds, surfaced to the user |
| AMBIGUOUS | more or fewer faces than declared | blocks meshing |
| BROKEN | matches no face at all | blocks meshing |
A named selection is a name plus a rule, built from five clause kinds: position plane, position range, normal direction, area range and length range. Rules are data, never code, re-resolved after every edit; a blocking verdict stops work at creation, at meshing and at hand-off to simulation.
h1 = 2 · y⁺ · ν / uτ · uτ = √(τw/ρ) · τw = (f/8)·ρ·U² · f = (0.790 ln Re − 1.64)−2
The engineer states a target y⁺, cells across a reference length, a resolution rung and the wall roles; the first-cell height is derived, with y⁺ quoted at the cell centre. The friction factor follows Petukhov's correlation and warns outside its fitted range. The refinement ladder varies one scalar by a ratio of 1.5 and holds the near-wall cell fixed; the mesh verification chapter takes it from there.
SICN = 3 / (‖A‖F · ‖A−1‖F) · 1 regular, < 0 inverted · floor 0.02
A design one small step from a solved one can reuse that mesh, moved to the new wall with the volume following by elasticity, the path on which adjoint gradients are exact. Every moved mesh is judged by the signed inverse condition number of each tetrahedron, no element inverted and the worst above the contract's floor. A step that fails is remeshed from its CAD and solved, never rejected, and its record says so. The adjoint optimisation chapter carries gradients through the moving mesh.
The S-duct is the reference family: four design variables, a pressure-drop objective, a half model on a symmetry plane. Every value below was measured and read from the study's own records; the acceptance bars were declared in code before the family existed, and none was moved.
The search inside this box returned a design with 8.0 % lower pressure drop than the baseline (79.30 → 72.98 Pa), verified by the solver at one operating point, 10 m/s, and 8.9 % lower again on a second, independently generated mesh. Measured The more useful finding: three variables pressed against their bounds, so the next question is a wider box. The same pipeline drives a finite-element solver: an L-bracket and a heated-bracket family each solved 65 of 65 designs. Measured
A deliberately large step, halfway to the far corner of the box, left the moved mesh with 228 inverted elements. The design was rebuilt from CAD, meshed afresh and solved to a converged 118.96 Pa; its record says so. Measured
Reviews start from the space: which variables matter, which bounds bind, which variants could not be built, before optimisation budget is spent.
No re-exporting, re-tagging or re-meshing by hand. Boundary conditions follow named faces through every rebuild, and a variant that breaks them stops before it reaches the solver.
A campaign is a planned, counted, resumable job with a fixed seed. What was asked, answered and left out is on record, ready for provenance and audit.
| Step | Typical workflow | With TensorLynx |
|---|---|---|
| Making a variant | Edit the CAD, export, re-import | Regenerated in memory from bounded variables |
| Boundary names | Re-tagged by hand, or by face number | Resolved by rule; ambiguous or broken blocks the mesh |
| Meshing | Mesher settings copied between variants | Each design meshed from stated intent, in an isolated worker |
| Sampling | A spreadsheet of chosen points | Seeded Latin hypercube, Sobol' or factorial; Sobol' balance checked |
| Failures | Dropped or overwritten | Recorded with stage and reason; missing is never zero |
| Hand-over | A folder of files | A campaign that opens like solver output, and STEP read back |
A mapped space is what the rest of the loop learns from: trust-aware surrogates are trained on these campaigns, and the TensorLynx overview shows where each chapter takes it next.
Both. STEP, IGES and BREP files are imported with shape healing and a free-edge report. To explore a part, the dimensions you want to vary become bounded design variables driving an editable feature tree; on a first family our engineers usually set this up with you.
Latin hypercube, Sobol' and full factorial, drawn in the normalised unit box. The seed is fixed by default because a plan whose points change between runs cannot be reproduced. A Sobol' count that is not a power of two is flagged with both neighbouring counts.
It is recorded with the stage and the reason, never dropped and never filled with a number. A solve that has not settled gets one longer run; if it still has not settled it is left out of every model, with its reason shown.
Each variant is meshed on its own from engineering intent: a target y+ and cells across a reference length. Boundary names follow faces by rule and are cross-checked by face area. A moved mesh that fails its quality check is replaced by a fresh mesh from CAD.
No. Exploration maps the space before an optimisation budget is spent on it: which variables matter, where the response is steep, which bounds bind. The same campaign then feeds the rest of the loop, from coupled physics to surrogates and multidisciplinary optimisation.
The campaign, laid out as solver output so every design opens in Results Analysis and History; the record of what was planned, solved and left out; and the chosen design as CAD, exported as STEP and read back.
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.
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.
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.
Our engineers run the loop on your problem: physics, sensitivities, optimisation and full-model verification, handed over with every setting on record.
A TensorLynx workflow, surrogate and verification gates set up around your product family, so the next design cycle starts where the last one ended.
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.
