Physics-informed models for engineering simulation

Real-time predictions from the simulations you already have.

Invariant Physics turns tens of COMSOL or Ansys runs into a model that answers the next design question in milliseconds, and tells you how far to trust the answer.

9 training runs38 ms per new caseRuns on a laptop CPU

The problem

Every new design question can mean another simulation.

  • Change the load, the flow rate or the material, and engineers often run the solver again. A single run can take hours; a study can need hundreds.

  • Most teams have tens of runs per question, not thousands, and they are rarely tidy: different meshes, settings and units across projects.

  • The runs are design IP, often under customer NDA or export control. Training on a vendor's cloud means uploading them.

  • And an engineer has to sign off on the number, so a prediction without a tested error isn't much use.

Force 400 → 1200 N Load angle 0° → 90°
Nine COMSOL runs of one steel bracket: surface von Mises stress on a shared colour scale.

How it works

Learn more from fewer runs, and test every answer.

Neural surrogates work well when you have hundreds of runs and GPUs. Invariant Physics is built for the tens of runs most teams actually have.

  1. Import the runs you have

    COMSOL or Ansys exports, experiments, parameter sweeps, or CSV and Excel tables. Each run stays on its own mesh.

  2. Find the physics they share

    It reads the units, builds the dimensionless groups, and looks for scaling, symmetries and relationships between fields.

  3. Test before you trust

    It holds back runs, predicts them, and reports the error field by field. Cases outside the tested range are flagged.

COMSOLfull solve
Von Mises stress on the bracket computed by COMSOL
Peak von Mises
52.6 MPa
Invariant Physics38 ms on a CPU
Von Mises stress on the bracket predicted by Invariant Physics
Peak von Mises
52.6 MPa

Load angle 67.5°, force 1000 N: a case the model never saw. Trained on the nine runs above. Same colour scale.

Demo

Nine bracket runs in, a new load case out.

Screen recording on a CPU. Training sped up as marked; prediction in real time.

What's different

Structure you can read, not just weights.

AspectTypical neural surrogateInvariant Physics
What it learnsWeights that map inputs to fieldsDimensionless groups, symmetries, equations and laws you can read
Runs neededTypically 30 to several hundredTens: 9 for the bracket, 18 for a wafer film
Training hardwareGPUsA laptop CPU
Outside the training rangeAccuracy usually falls offFollows the law it found where one holds, and flags every case outside
What comes with an answerA numberThe method used, the error it was tested at, and where it holds
Your runs stay with you.

Training happens on an ordinary CPU, so simulation data never has to leave the company. No GPU cluster, no upload.

Results

Measured on cases it never saw.

30/30

unseen airfoils predicted closer than the nearest simulation

AirfRANS public dataset, velocity
0.05%

error 10× beyond the training range, against 35% for standard interpolation

Controlled boundary-layer benchmark
15

public CFD and heat-transfer datasets, same default settings

Plus 20 COMSOL studies
Velocity profile chart: Invariant Physics follows the exact boundary-layer solution 10 times beyond the training range, while standard interpolation is off by 35 percent
Boundary layer trained on x ≤ 0.3 m, predicted 10× farther along the plate.

Tested across the physics engineers simulate

  • FluidsCylinder wake, airfoils, pump, static mixer, turbulent channel
  • Heat transferHeat sink, heat exchanger, turbine stator, battery pack
  • StructuresLoaded bracket, piezoelectric micropump, elastoplastic plate
  • ElectromagneticsPermanent-magnet motor, eddy currents, induction heating
  • PlasmaReference cell, argon discharge
  • ChemistryTubular reactor, Li-ion battery cell

Tested is not solved. Accuracy varies by problem, and some cases still fail. We report those too.

Who it's for

Teams where a wrong answer means a recall or a failed certification.

  • SemiconductorsPlasma etch and deposition, thin films, chip and package heat
  • Automotive and EVStructural loads, aerodynamics, electric motors, battery cooling
  • Aerospace and defenceExternal aerodynamics, turbine cooling, brackets and frames
  • Energy and processHeat exchangers, pumps, mixers, reactors
  • ElectronicsThermal management, MEMS and micro-pumps, induction

Early access

We're working with a small number of engineering teams.

If you have a design question and a folder of simulation runs, we'd like to hear about it. Tell us what you simulate, which solver you use, and roughly how many runs you have.

Email contact@invariantphysics.com
Open in mail app