Simulating a power kite once took hours of supercomputer time and now a machine learning model does it in the blink of an eye
Running a full aerodynamic simulation of a power kite once demanded nearly 20,000 seconds of CPU time. A new computational framework does the same job in about 25 milliseconds.
The kites in question aren’t ordinary aircraft. Leading-edge inflatable kites are soft, swept-back membrane wings — their tubular leading edge pre-inflated like a bicycle tire, their canopy held taut by tension alone — designed to harvest wind energy or pull cargo ships across oceans. Their curved, low-slung shape makes them extraordinarily difficult to model: the inflated tube triggers flow separation on the pressure side that conventional aerodynamic solvers can’t handle.
Why kite aerodynamics is so hard to model
Power kites break almost every assumption that standard aerodynamic tools rely on. Their swept-back planforms, low aspect ratios, and pronounced anhedral — the downward droop of the wing tips — combine to produce flow patterns that classical methods weren’t built for.
The team validated the framework against wind tunnel measurements of the TU Delft V3 kite — a rigidized subscale model tested with a six-component load balance.
The inflated leading-edge tube triggers persistent flow separation on the pressure side across the entire flight envelope. Solvers like XFOIL, which assume a thin boundary layer mostly attached to the surface, lose physical validity the moment that separation takes hold. For LEI kites, that’s essentially always.
The high-fidelity alternative — full 3D Reynolds-averaged Navier–Stokes simulations — is accurate but punishing. A single run of the TU Delft V3 reference kite required roughly 19,000 seconds of CPU time, which rules out iterative design work almost entirely.
A hybrid solver built for speed and fidelity
The research team paired a computationally lean aerodynamic solver with high-quality airfoil data. Their chosen solver is the vortex step method (VSM), a lifting-line approach that captures three-dimensional induction effects efficiently. Rather than classical potential-flow polar data, the team supplied polars derived from 2D RANS simulations — capturing the viscous separation behavior that defines LEI airfoils.
Building that polar database required serious upfront investment. More than 100,000 RANS CFD simulations were run on a university cluster, covering a wide range of airfoil shapes, angles of attack, and Reynolds numbers, totaling roughly 140,000 core hours. The payoff: a Julia implementation of the VSM evaluates a 150-panel kite in approximately 25 milliseconds — about one million times faster than the 19,000-second 3D RANS baseline.
Machine learning fills the gap between simulations
Running 100,000 CFD simulations is expensive once. Running them every time you tweak a design isn’t feasible. The team trained an extremely randomized trees (ERT) surrogate model on the RANS database to predict lift, drag, and moment coefficients for any parameterized LEI airfoil shape — give it geometric parameters and an angle of attack, and it returns aerodynamic coefficients instantly.
The model achieves an R² above 0.98 at all three Reynolds numbers tested: 1×10⁶, 5×10⁶, and 2×10⁷. The authors note this is the first application of a machine learning surrogate trained on RANS data specifically for LEI kite aerodynamics. One practical refinement: tree-based models produce piecewise-constant outputs, which can cause problems in optimization loops requiring smooth derivatives. The team addressed this by fitting outputs with piecewise cubic Hermite splines, producing continuously differentiable polars without spurious oscillations.
Validation: beating 3D CFD with a fraction of the cost
The team validated the framework against wind tunnel measurements of the TU Delft V3 kite — a rigidized subscale model tested with a six-component load balance. Within the operational angle-of-attack range, 3D lift predictions agreed with measurements to within 9%, and drag to within 13%.
The VSM-plus-ML framework reproduced measured aerodynamic trends more consistently than full 3D RANS simulations, which carry a geometric offset and are sensitive to whether struts are included. At the 2D airfoil level, the ML model’s polars agreed with wind tunnel data within 4% in lift and 16% in drag in the near-linear regime. The framework also correctly predicted static stability in roll, pitch, and yaw for the V3 kite’s trim condition, along with aerodynamic damping in all three axes.
Design trade-offs revealed by parametric studies
With a fast, validated tool in hand, the team ran parametric studies to map the design landscape for LEI kites. Varying airfoil camber height, leading-edge tube diameter, maximum-camber position, and trailing-edge reflex angle revealed a consistent tension: configurations that maximize lift-to-drag ratio don’t align with those that maintain healthy stability margins.
Increasing wing aspect ratio raises aerodynamic efficiency and strengthens yaw-restoring moments, while greater anhedral improves turning ability at the cost of efficiency. Tow-point position strongly affects trim angle and pitch stability — place it too far aft and yaw instability follows. There’s also a counterintuitive finding: for kites rotating about a low tow point, anhedral is stabilizing in roll, the opposite of what conventional aircraft experience with dihedral.
What comes next for airborne wind energy
The validated framework is openly available, and the TU Delft V3 kite is positioned as a community benchmark for future aerodynamic research. The immediate next step is aero-structural coupling — updating spanwise airfoil polars in real time as the flexible wing deforms under load. A kite in flight isn’t a rigid body, and the gap between rigid-wing predictions and real deformed-wing behavior is where the next round of modeling errors likely lives.
Incorporating bridle geometry and tow-point optimization into the design loop would allow simultaneous tuning of performance and stability. A framework that evaluates a full kite in 25 milliseconds opens doors that were previously closed: real-time flight control, large-scale design optimization, and the rapid iteration that commercial airborne wind energy will require.
Reference: Poland, J. A. W., Masure, K. R. G., Cayon, O., and Schmehl, R.: Computational aerodynamics for leading-edge inflatable kite design, Wind Energ. Sci., 11, 4147–4178, https://doi.org/10.5194/wes-11-4147-2026, 2026.
Daniel García is an Editor-in-Chief with strong expertise in structural work and engineering principles. He combines this technical foundation with deep knowledge of energy, spatial design, and emerging technologies, bringing a forward-thinking and analytical approach to editorial leadership.
