HeatSink Sizer
Heat sinks in seconds, checked against published correlations.
0–5% vs. literature
Use cases
What we have measured, against what, and where we are not there yet.
Every row below has a reference external to training: a published experiment, a finite-volume or finite-element solver, or an industry baseline. Where there is a limit, it is stated beside the number.
Evidence
Technical validations
| Case | Metric | Reference | Status |
|---|---|---|---|
| Thermal surrogate7 heat-sink geometries | ~200,000× · 4.46% | 3D FEniCS/dolfinx + gmsh; worst case 7.35% | validated |
| Erosion in a 90° bendOpenFOAM + surrogate | 1.05–1.49× · 1.0% | Vieira et al. 2016, open experiment; factor-2 criterion | validated |
| Airfoil optimizationsurrogate-guided, re-verified | −18.2% | Profile drag; target lift held to 0.04% | validated |
| Vehicle aerodynamics10 simpleFoam runs | 0.05% · 1.3% | Velocity field and coefficients vs. OpenFOAM; Cd at 0.17% | validated |
| PINN vs. finite volumesfield-by-field comparison | 0.39% | Independent finite-volume solver; max 0.67 °C | validated |
| Lid-driven cavitymesh refinement | 0.19% | Ghia et al.; starting from 1.26% | validated |
| Inverse h with a PINN2D plate and 3D block | 14.8 ± 0.3 | True value 15; hot spot 79.2 °C vs. 79.1 °C | validated |
| PCB thermodynamics8 package families | within 8% | Published θJA, from SOIC-8 to a 35 mm BGA | validated |
| Rover mobilitypublished model, corrected | 0.7–3.8% | Bekker–Wong; traction error 45× lower than the original | validated |
| Brake-disc Monte Carlomanufacturing tolerance | 1.9 °C | Nominal 495.9 °C; 95th percentile at 498.1 °C, limit 500 °C | validated |
| HX-201 twinfouled exchanger | 0.216 °C · 95.5% | Redundant probe kept out of the filter; true 0.001637 m²K/W | validated |
| Physics confidenceclient model, confidential | residual >100× | VeriPhysics guardrail over an already-trained model | validated |
Figures
The pictures behind the numbers
And more
Built on the same base
The platform holds 63 applications. These are the ones behind the four products above.
PCB Hotspot
Junction temperature of every part in about a second.
θJA within 8%
Simulation Preflight
PASS, WARNING or FAIL before the run, with file and line.
caught before the run
Mesh Quality
OpenFOAM, Gmsh, VTK, Abaqus, STL: metrics, heatmap, worst elements.
identical to checkMesh
Inverse Heat Lab
The h coefficient from a few thermocouples, with a PINN.
h = 14.8 ± 0.3
Aircraft Design Optimizer
NSGA-II over the wing, checked in 3D OpenFOAM.
−13% CO₂/pax-km
Surrogate Studio
Simulation results become a fast model with enforced validity.
versioned endpoint
Twin Runtime
A validated model beside operational data, with EnKF and revalidation.
MQTT · REST · OPC UA
Evidence Report
Experiments become reproducible evidence with SHA-256 on everything.
only what ran passes
Where we are not there yet: Erosion relies on a factor-2 band, not 1% precision. Flat residuals do not prove drag has settled. And in one cost analysis a classical numerical method was faster and more accurate than the network itself — the recommendation was the classical method. The twin's band is a state band calibrated on a separate window, and its lead comes from a linear extrapolation whose error is reported.
Bring your case
Got a simulation that is slow, expensive or hard to verify?
Send it over. We come back with a closed scope and a measurable success criterion — or we tell you the classical method does it better.