Wing Optimizer
From a set of mission requirements to a shortlist of wings worth investigating.
Too many combinations.
A limited solver budget.
Span, area, taper, sweep, washout, airfoil choice, and CG all influence a wing’s performance. Testing every combination with a three-dimensional aerodynamic solver quickly becomes impractical.
Wing Optimizer manages that search: it screens section data, proposes geometries, starts fresh Flow5 evaluations, checks the returned evidence, and keeps a record of the candidates. A learned model guides future trials; saved solver results determine the reported performance.
Real saved flying-wing searches from 15 and 17 October 2026. Both use a 0.7 kg mission at 10, 20, and 30 m/s. The later tailed-main-wing searches are excluded from these results.
A search loop with
evidence at every step.
The current application combines model-guided proposals with independent exploration and direct refinement.
- 01
Define the mission
Mass, speeds, geometry limits, airfoil pools, CG bounds, and flight-role requirements.
- 02
Screen the airfoils
Use compatible section polars to build root and tip shortlists.
- 03
Propose a wing
Balance learned suggestions, new geometry, direct local moves, and foil-pair challenges.
- 04
Evaluate in Flow5
Solve fresh fixed-control polars at the mission speeds, then fine-check promising candidates.
- 05
Check and rank
Verify evidence, required lift, drag, and active constraints. Apply the extra drag allowance once.
- ↺
Learn and repeat
Use completed observations to select the next useful trials until the budget ends.
- 06
Export and review
Save the shortlist, charts, Excel comparisons, checkpoints, and available native Flow5 projects.
The scheduler is custom work inspired by local and constrained Bayesian optimization. It is not the reference implementation of TuRBO or SCBO.
Explore the finalists.
Choose a run, then select a candidate. The figures below compare the two saved winners; the table exposes every saved finalist in each run.
Selected candidate
| Point | Speed (m/s) | L/D | Drag (N) | Cm | α (°) |
|---|
| Rank / select | Root / tip | Span (m) | Mission L/D | Cruise L/D |
|---|
Two runs, shown side by side.
Mission L/D · JX-GS-15 / MH30
Cruise L/D 8.59 · Cm −0.0317
Mission L/D · E171 / RG15
Cruise L/D 7.96 · Cm −0.0242
The mission mass and speeds match, but other requirements and scoring settings changed. These are recorded runs, not a controlled benchmark of algorithm improvement.
Mission L/D includes the declared extra drag coefficient of 0.012. Mission weights are 25% slow, 55% cruise, and 20% fast. The displayed winners have saved applied-settings, viscous-drag, and numerical-agreement checks.
Physical elevon trim and flight readiness are not verified. Both searches ended at a 14-hour limit; neither establishes a global optimum or general aerodynamic accuracy bound. Raw wing-only polars can show higher L/D because they omit the extra drag allowance.
Built, checked,
and revised.
The case study combines the original wing-optimization discussion, the local Codex development record, and the saved run files. The public summary below focuses on the engineering decisions.
Move beyond a fixed shortlist
Learn from completed wing evaluations while keeping independent exploration, foil challenges, and direct local refinement active.
Keep forecasts separate from results
A predicted score cannot certify a candidate. Saved coefficients, input identities, qualification checks, and coarse/fine comparisons support the result.
Make long runs recoverable
Dated runs, checkpoints, preserved settings, and native-project exports make it possible to inspect and recover work later.
Fix Reynolds coverage screening
Respect feasible geometry when screening the library, and let the user fit the Reynolds grid to larger aircraft. Historical results above remain the original saved October runs.
Data provenance and references
Results are extracted from the saved summary.json files for runs 2026-10-15_22-45-34 and 2026-10-17_00-33-15. Geometry images also use their exported airfoil coordinates. No simulations were run for this page.
The downloadable JSON includes the source-summary hashes, mission assumptions, saved candidate coefficients, and polars. Local paths and private chat transcripts are excluded.