ML surrogate · upstream of CFD
Screen an airfoil in 1.9 ms —
before you spend hours in CFD.
AirfoilLearner predicts Cl, Cd and Cm for any 2D section in real time. Fail fast on the weak geometries and save the cluster for the designs worth simulating. It sits upstream of your solver. It doesn't replace it.
80+ aerospace engineers interviewed · 60+ on the waitlist
trained on 50,000+ CFD simulations · UIUC airfoil database
The problem
High-fidelity CFD is the bottleneck in early aerodynamic design. It's slow, it's expensive, and one wrong setup can burn a day of compute — so engineers test a handful of shapes when they should be testing thousands.
£75k–£2.5M
what a single 500-run 3D RANS design campaign costs in compute.
30–40%
of an engineer's time goes to re-running simulations after a bad setup.
a few dozen
geometries actually get evaluated, out of the thousands worth trying.
How it works
A screening layer, not a CFD replacement.
Most "AI for CFD" tools are a faster XFOIL with a nicer interface. AirfoilLearner is the whole funnel: it decides which designs deserve expensive compute, and hands the rest a fast, directionally-honest answer.
01
XFOIL sweep
Every candidate geometry gets a fast 2D pass in under 5 seconds.
02
BNN uncertainty
A Bayesian net scores where the surrogate is unsure and flags bad training data.
03
RANS escalation
Only the top 10–20 candidates are escalated to expensive higher-fidelity CFD.
04
Surrogate + FlowSense
Real-time Cl, Cd, Cm plus a plain-language read on which parameter to move.
05
OpenFOAM handoff
The shortlist drops cleanly into your 3D solver — warm-started, not cold.
The numbers
Fast where it can be. Careful where it counts.
1.9 ms
to predict Cl, Cd and Cm for a 2D section
NACA 4412 · Re 2.0×10⁶
9.5–14.8%
of candidates ever needed full RANS in published runs
the rest handled by the surrogate
50–70%
less training data via active sampling
vs uniform sampling of the design space
50,000+
CFD simulations behind the model
UIUC database + generated data
AirfoilLearner is at research stage. These figures come from our pipeline and the published literature it's built on — not marketing rounding.
The science
Why it holds up.
The hard part isn't predicting lift quickly. It's knowing when a fast prediction is safe to trust — and routing the rest to the solver before they cost you a week.
2D sections today → 3D wings → full-aircraft geometry next.
Beyond XFOIL's limits
XFOIL is fast and free, but it breaks down in separated flow and at high angles of attack. The surrogate is trained to stay honest exactly where the cheap tools stop being trustworthy.
It knows what it doesn't know
A Bayesian neural network attaches an uncertainty band to every prediction and decides when a design is worth escalating to higher-fidelity CFD. The triage is automatic, not a manual judgement call.
From insight, not just numbers
The FlowSense layer reads adjoint sensitivities and tells you which geometric parameter to move and why — so you get a design direction, not just a coefficient.
Try it now
Run your first section before you read the docs.
No meshing, no licence, no setup. Pick a profile, set the angle, watch the coefficients land in milliseconds. The demo is open and free.