27% more lift for a wing that flies close to the ground or water.

Ground-effect craft skim just above the surface on a cushion of air, and run short of lift during take-off. Archimedes designed a wing section with 27% more lift under the same drag limit, in 2D simulation.

Talk to us Download report PDF

Recorded geometry

Recorded CAD boundary tessellation of the exact 2D sections with arbitrary display-span extrusion. Source dcbe93af0e3a90a3. Drag to inspect.

27%more lift than the best standard wing section, same drag limit

$174total cost of compute and model use for the run

938designs tried on 32 cloud machines in 3 hours

The ask

Get the most lift from a wing section flying low, without passing a drag limit.

Any number of wing elements allowed, inside a fixed box.

What it found

Tilting a single wing harder stalls it.

Past about 11 to 13 degrees the airflow separates and drag shoots past the limit.

What it did

Split the wing into a main element and a slotted flap.

Air through the slot keeps the flap working, while the angled flap builds more pressure under the wing.

Fine print

More lift, not more efficiency.

The standard section still has the better lift-to-drag ratio, and stability in pitch hasn't been checked.

Got a shape that needs more from the same limits?

Talk to us See all cases

For engineers

The full record, unedited.

Interactive comparisonexplore the measured results

Measured comparison

Lift under one fixed drag ceiling

All candidates use the same moving-ground RANS evaluator at h/c 0.15. The plot compares lift while the plate holds the exact S1223 reference and feasible W2b geometry at one fixed projection. Lift and drag must be read together because the exact winner is mesh-sensitive.

Recorded 2D sections at h/c = 0.15. 1.0 m reference chord. Geometry is reconstructed from the recorded campaign artifacts.Geometry only. No inferred streamlines, pressure field, or unrecorded physical detail is shown.A source: results/evolve/robustness/wig_baseline/runs.jsonl · 2990ef2454ceb6c2 · Recorded CAD boundary tessellation of the exact 2D section with arbitrary display-span extrusion.
B source: results/evolve/robustness/wig_repeat/runs.jsonl · dcbe93af0e3a90a3 · Recorded CAD boundary tessellation of the exact 2D sections with arbitrary display-span extrusion.

Lift coefficient · Clhigher is better →

Drag coefficient · Cd← lower is better · cap 0.030

Drag is plotted only where the case record reports a paired value. The nearby W2b point remains below the Cd 0.030 acceptance line.

Numerical basis: 2D steady RANS (OpenFOAM simpleFoam, k-omega SST, moving ground wall, 1.5x refined mesh)

Lift at the drag cap

Moving-ground RANS at h/c = 0.15 and Re 3 million. Drag is shown only where the record gives a paired value.

Section	Cl	Cd	Status
NACA 4-digit optimum	1.707	At cap	Single-element separation limit
NACA 5-digit optimum	1.737	At cap	Single-element separation limit
Selig S1223 optimum	1.829	0.0196	Best conventional L/D, 93
Nearby feasible W2b	2.331	0.027887	Drag-cap-compliant search point
Refined W2b candidate	2.37522	0.028852	One refinement, not a GCI study

Why the exact repeat is not the headline

The exact search-mesh repeat crosses the acceptance line, so the reported lift gain uses the nearby feasible point.

Solve	Cd	Against cap 0.030
Nearby feasible point	0.027887	Accepted
Exact search-mesh repeat	0.04365	Rejected by the drag cap
Refined candidate	0.028852	Mesh-sensitive, one refinement only

Applicability

The +27% lift claim belongs to the nearby feasible 2D point, not to a grid-converged exact winner. Pitching stability, span effects, and a free surface are unresolved. S1223 retains the higher lift-to-drag ratio.

References and source records

Selig S1223 high-lift section, used as the optimized single-element reference.
Campaign lock: Re 3 million, 10 m/s, moving ground, k-omega SST, h/c 0.15, Cd at or below 0.030.

The exact briefas given to the agents

Maximize section lift coefficient Cl inside a 1.0 m x 0.5 m design box at ride height h/c = 0.15 above a moving ground wall (Re = 3 x 10^6, k-omega SST) subject to drag cap Cd <= 0.030. Open topology: any number of airfoil elements with >= 8 mm clearance.

Archimedes: 3.0 h autonomous swarm search (938 CFD solves)
Compute record: 32 cloud workspaces · $174 total compute + tokens

Decision loghow the agents reasoned, step by step

Reasoning trace

What changed, and why

The search first optimized the single-element references, then coordinated slot width, overlap, and flap angle in the literature-informed multi-element families so pressure capture and boundary-layer attachment improved together.

Literature-Informed Topology Search

Co-optimizing a slotted section under a hard lift and drag objective

Observed: Single-element sections pushed above Cl ~ 1.83 separate on the rear upper surface or choke the ground channel, violating Cd <= 0.030.
Inferred: Splitting the section into a main element and a steeply deflected trailing flap closes the exit nozzle toward the moving ground (raising cushion Cp), while a converging 21 mm slot bleeds high-pressure under-wing air onto the flap's upper boundary layer.
Changed and measured: The refined candidate reached Cl = 2.37522 at Cd = 0.028852. The exact geometry exceeded the drag cap when repeated on the search mesh (Cd = 0.04365), while a nearby search point remained feasible at Cl = 2.331 and Cd = 0.027887.

Engineering findingheadline numbers and the mechanism

More useful lift from the same geometric box

For a ground-effect craft, the computed lift margin could become payload or lower takeoff speed. The present result is a promising 2D section candidate, not yet a stable vehicle design.

+27.4%lift coefficient over the best conventional single-element section under the same Cd <= 0.030 drag cap (Cl 2.331 vs 1.829 Selig S1223)2D RANS CFD · feasible search-mesh point · h/c 0.15

0.0279drag coefficient of the nearby feasible slotted two-element point (inside the 0.0300 budget)Search mesh · Re 3e6

2.375 / 0.0289lift and drag of the refined candidate; its exact geometry repeated at Cd 0.04365 on the search mesh1.5x refined mesh · mesh sensitivity disclosed

Why the mechanism works

Among the literature-informed topology seeds, the strongest candidate combined a 5.1%-cambered main element at 9.6 deg incidence with a 26.6 deg aft flap and a 21 mm converging slot. The computed slot flow re-energizes the flap boundary layer while the deflected element raises ground-cushion pressure.

What the baseline got wrong

A two-seat ground-effect craft (ekranoplan) is lift-limited during low-speed take-off over water. Pitching a single-element airfoil above 11-13 deg to trap higher ram pressure under the wing causes massive trailing-edge flow separation that explodes drag past the Cd <= 0.030 limit.

Full technical recordappendices, families and limits

Campaign record and limitations

Appendix

The detailed brief, decision record, search families, solver basis, and verification checks are exposed below without tabs or collapsed panels.

Appendix A: Frozen brief and campaign scale

Maximize section lift coefficient Cl inside a 1.0 m x 0.5 m design box at ride height h/c = 0.15 above a moving ground wall (Re = 3 x 10^6, k-omega SST) subject to drag cap Cd <= 0.030. Open topology: any number of airfoil elements with >= 8 mm clearance.

Search duration: 3.0 h
Evaluations: 938
Compute fleet: 32 cloud workspaces
Solver basis: 2D steady RANS (OpenFOAM simpleFoam, k-omega SST, moving ground wall, 1.5x refined mesh)

Appendix B: Full decision log

01

Hour 00:00 - 03:00 · Literature-Informed Topology Search

Co-optimizing a slotted section under a hard lift and drag objective

Observation: Single-element sections pushed above Cl ~ 1.83 separate on the rear upper surface or choke the ground channel, violating Cd <= 0.030.
Reasoning: Splitting the section into a main element and a steeply deflected trailing flap closes the exit nozzle toward the moving ground (raising cushion Cp), while a converging 21 mm slot bleeds high-pressure under-wing air onto the flap's upper boundary layer.
Action and result: The refined candidate reached Cl = 2.37522 at Cd = 0.028852. The exact geometry exceeded the drag cap when repeated on the search mesh (Cd = 0.04365), while a nearby search point remained feasible at Cl = 2.331 and Cd = 0.027887.

Appendix C: Search families

37 section parameterizations. These are distinct programs or engineering configurations, not cosmetic renders.

Cambered main element + 26.6 deg slotted flap (feasible W2b point, Cl 2.331 at Cd 0.0279)
Slotted flap + Gurney tab family (Run W1, Cl 2.21)
Bayesian-optimised Selig S1223 high-lift baseline (Cl 1.829 at Cd 0.0196)
Bayesian-optimised NACA 4-digit (Cl 1.707) & 5-digit (Cl 1.737) families

Appendix D: Numerical basis and verification

2D steady RANS (OpenFOAM simpleFoam, k-omega SST, moving ground wall, 1.5x refined mesh)

Fair baseline: the swarm's own Gaussian-process Bayesian optimizer searched NACA 4-digit (44 evals), NACA 5-digit (34 evals), and Selig S1223 (13 evals) up to 16 deg incidence under the identical evaluator and drag cap.
Winner and S1223 baseline were re-solved back-to-back on one 1.5x finer body-fitted mesh. This is a refinement check, not a grid-convergence study.
The optimized S1223 has the better L/D (93 versus 82). The winner's more negative pitching moment, 3D tip effects, free-surface coupling, and dynamic stability remain unresolved.

Appendix E: Reported outcomes and limits

+27.4%: lift coefficient over the best conventional single-element section under the same Cd <= 0.030 drag cap (Cl 2.331 vs 1.829 Selig S1223) Basis: 2D RANS CFD · feasible search-mesh point · h/c 0.15.
0.0279: drag coefficient of the nearby feasible slotted two-element point (inside the 0.0300 budget) Basis: Search mesh · Re 3e6.
2.375 / 0.0289: lift and drag of the refined candidate; its exact geometry repeated at Cd 0.04365 on the search mesh Basis: 1.5x refined mesh · mesh sensitivity disclosed.

Explicit limitations

Only one body-cell refinement was run; this is not a grid-convergence study.
The exact winner repeats above the drag cap on the search mesh, while a nearby point remains feasible.
S1223 has the better L/D (93 versus 82), and pitching stability, 3D tips, free-surface effects, and waves remain unresolved.

Recorded Archimedes run: 3.0 h autonomous swarm search (938 CFD solves). Compute record: 32 cloud workspaces · $174 total compute + tokens.