Wing Optimizer 2.6.1 release notes

Version 2.6.1

Correct the Reynolds envelope used for cached-airfoil screening: span/area corners excluded by aspect ratio or minimum tip Reynolds no longer force unnecessary library coverage. The same minimum-Re limit now filters screening scenarios. Foil-specific thickness and lift coverage still need to pass.

Add Fit Reynolds grid to aircraft in Airfoil library. It preserves current levels and adds endpoint levels in 1.5x steps, including above 3 million when the aircraft requires them. It edits the displayed grid only. Preview and manual generation still use exactly the selected levels. Missing verified coverage reports both required and available Reynolds ranges.

No saved settings, real polar data, run records or aerodynamic scoring rules are changed. See the guide and validation.

Version 2.6.0

Fix CG-sensitive diversity and fine-promotion identities for balanced flying-wing scoring when optional moment objectives are off. A fresh mocked end-to-end case shows a previously suppressed better CG receiving an actual fine check.

Reject materially negative reported integrated induced drag at the operating stencil, independent of optional viscous columns or stored proof availability. Missing evidence remains unverified. Tailed-main-wing descriptions now state that aircraft pitch objectives/gates are inactive, without changing settings.

The research audit reviews primary optimization papers, five measured configurations and the latest three completed 3D runs. An alternative regional sampler was tested and rejected after mixed results included material quality regressions. The current sampler and aerodynamic score are retained. See the guide and validation. No real solver was started.

Version 2.5.0

Bounded shared training now retains strong balanced-objective observations, checked coarse sources and efficiency/moment alternatives alongside exploration and diversity. Incremental distance calculations and proposal-scoped evidence reuse reduce model-data processing overhead without caching solved wing setups.

Nonlinear penalty uncertainty follows folded/rectified Gaussian marginals. Fine-promotion forecasts and incumbents use the same raw utility basis, avoiding comparison with a final score containing crossing evidence the forecast cannot claim. Invalid nonfinite/negative uncertainty and lift/drag evidence are rejected. Measured scoring and hard qualification requirements remain unchanged.

Logs/debug exports separate shared and local training counts and prediction tasks. Compare saved settings… identifies material differences between saved runs without solver work. Ordinary GUI/CLI Excel exports publish under unused names atomically; an optional Flow5 handoff failure does not hide a saved workbook. Explicit CLI replacement requires --overwrite.

See the 2.5 guide and validation record. Tests use saved fixtures, mocked solvers and fictional landscapes. No real solver or library batch starts during this update, and no universal search improvement or physical-accuracy claim is made.

Version 2.4.0

The late search reserve now includes explicit model-guided and direct local refinement, with foil and global challenges retained. Fine promotion uses recorded rough/fine discrepancies to defer clearly uncompetitive old checks, protects efficiency and moment alternatives, and retains bounded independent coverage and near-boundary rescue. A near-cap guard prioritizes credible late challengers for fresh fine evidence. A scheduling deferral is not rejection or convergence; physical and numerical qualification checks are unchanged.

New completion logs and saved statistics include measured invocation and resume time, solver/evaluation timing distributions, counts, stop reasons, best-result progress and evidence limits. Prospective solver calls have separate failure/timeout records. Hard cancellation marks only the exited child's invocation and leaves unknown internal timing unknown.

Results adds Export debug and View run summary. Debug workbooks retain all saved attempts and forecasts, model history, coefficient comparisons, pair/track exploration, geometry/CG, timing, settings and complete JSON evidence. New model snapshots include exact prior training/tuning IDs and available kernel parameters. Exports remain separate from ordinary results, use completed staging files and never overwrite a simultaneous export.

See UPDATE_2_4_GUIDE.md and the validation record. Tests use mocked solvers and saved fixtures; scheduling benchmarks use fictional aerodynamics and costs. No real solver check or physical-accuracy improvement is claimed.

Version 2.3.0

The surrogate scheduler can now propose geometry changes around complete, integrity-valid physical rejects before any wing passes every requirement. The shared categorical model remains available alongside pair models. Bounded RBF/Matérn kernel selection, conservative coarse/fine discrepancy estimates, numerically stable acquisition calculations and bounded runtime preferences guide proposals without replacing solver observations. Exploration, independent regions, alternative-foil challenges and direct polls remain active.

Fine promotion retains efficiency leaders and competing foil pairs, considers predicted fine utility and uncertainty, and records forecasts before observing their outcomes. Logs and result details distinguish trained models from executed model-selected trials. Prediction-error summaries include usable physical rejects; missing or integrity-invalid results remain unknown.

New configurations select search.pitch_objective_policy: balanced_trim_v2. Cruise |Cm| and a moderate static-margin target are soft preferences. The upper margin bound becomes a mission-wide excess-margin penalty, and a supported crossing near required cruise lift has a small bounded score preference. Required lift, restoring pitch slope and minimum positive static margin remain mandatory. A strict crossing gate remains an explicit option, disabled in the new example configuration. Missing-policy histories preserve legacy scoring.

Excel retains every original Compare column through BD and appends the policy and new costs. For the new policy, editable review L/D weights subtract the complete saved pitch penalty once; the solver score remains unchanged. Search evidence exposes acquisition mode, normalized constraint violation, estimated cost and model scope. Lower-memory, atomic export handling remains.

See UPDATE_2_3_GUIDE.md and the local validation record. Benchmarks use fictional aerodynamics and include regressions as well as improvements. No real solver or library batch starts for this update. Native multi-speed batching remains deferred until a user-started round trip can verify applied settings, viscous evidence and repeatability behavior.

Version 2.2.1

Large Excel exports use bounded formatting and compact saved-curve lookups to reduce spreadsheet dependency growth. Full result rows, solver polars, geometry, provenance and native comparison charts remain available.

Each export uses an isolated runtime job, streamed JSON and logs, and a staging workbook. The exporter validates its ZIP contents before replacing the destination atomically. A failed export preserves the previous workbook and recovery evidence; an open destination leaves the completed recovery workbook available to save.

This patch changes export handling only. It starts no solver and does not alter saved aircraft settings or aerodynamic scoring. See EXPORT_RECOVERY.md for usage and the completed local validation record.

Version 2.2.0

Automatic fine checks are bounded by immutable completed-attempt history. Unchanged missing-lift setups no longer repeat; interrupted jobs remain resumable. Persisted reserve decisions prevent fine-batch size from starving foil challenges. Unused finalist allowance returns to fresh search and changed finalists receive affordable diagnostic refreshes without endlessly repeating failed keys.

Neutral cruise mode replaces the negative-side moment preference with a symmetric required-lift residual. Real saved polars supply stable zero-crossing alpha, CL, L/D and alignment evidence. An optional strict 3D gate, cruise-angle target and static-margin preference are available. Tailed-wing mode and legacy histories retain their appropriate scope. Physical elevon trim remains manual.

Excel's static and dynamic defaults now select eligible saved leaders consistently. Missing requested polar samples remain gaps, full diagnostic provenance stays separate, and neutral assessment metadata is exported. Original solver values and the inspected run's original workbook are preserved.

See UPDATE_2_2_GUIDE.md and validation/update_2_2/validation_summary.json for the completed local checks. This update starts no real solver or library batch.

Version 2.1.0

New runs can use a short -4..12-degree rough sweep and finer refinement checks, with bounded fresh extensions when required lift or pitch derivatives need more support. Finalists receive supplementary -4..20-degree sweeps on both meshes within the same allowance. These diagnostics never replace refinement scores or become extra model training evidence.

Initial seeds and promotions preserve measured efficiency leaders, diverse physical shapes and competing foil pairs. Shared categorical warmup requires diverse usable observations and can start earlier. Alternative foil challenges continue during the refinement reserve, including a late challenger check before the diagnostic plan is frozen. Completed geometry results remain fresh solver evaluations; explicit repeats are supported.

The new optional cruise Cm preference acts at required cruise lift and can require Cm <= 0 for a flying wing. Static stability still uses restoring Cm slope and the selected margin band. Highest measured L/D references and an efficiency/moment frontier remain available. Physical elevon trim is manual. Existing runs retain their original pitch preferences and analysis mode on resume.

The preliminary GUI now combines refined foil-pair results at matching panel resolution and groups CG variants by physical wing in its default view. Full evidence remains accessible. Excel retains all source rows, adds stage and cruise moment metadata, compares review ranks within stages, and associates supplementary final polars with their source wing without changing its score.

See UPDATE_2_1_GUIDE.md and validation/update_2_1/validation_summary.json for the completed software checks. No real solver run was started for this update; the new staged solver workflow still needs a user-started round-trip check.

Version 2.0.0

V1 preserves the previous installed app and scientific data. V2 adds categorical surrogate warmup and separate foil-pair models, competing geometry regions, continuing exploration, direct-search fallback and bounded fine-mesh challenges for promising coarse constraint rejects. Predictions guide fresh solver trials; they never become aerodynamic observations or feasibility certificates.

V2 rechecks raw required-lift coefficients, common CG, extra drag, native applied settings and local viscous execution evidence. Known contradictions and failed required-lift viscous analyses reject candidates. Physical elevon trim stays manual. The GUI and complete exports are retained, with separate Search evidence and Untrimmed mission views and protected V1 destinations.

The local full suite passed 412 tests, and the final affected export/library suite passed 132 tests. Actual Excel checks passed 12 sheets, 20 native charts and 25 selected-cell checks. The approved six-job fresh Flow5 test completed with native settings and viscous proof and correctly rejected its unsuitable fixed wing. See V2_VALIDATION.md for scope; none of this certifies global or flight accuracy.

Version 0.3.1

Input fields no longer show hover popups. Action buttons now show short hover summaries covering what each action does and whether it starts solver work. The window title/header show 0.3.1. Added an offline process flowchart for the current coupled foil/geometry workflow. The search and scoring algorithms are unchanged from 0.3.0; aircraft settings and saved runs are preserved.

Hover regressions and the main/catalog GUI checks verify help on buttons and none on input widgets. See WORKFLOW_FLOWCHART.md.

Version 0.3.0

This update gives competing foil pairs their own geometry/CG refinement tracks, keeps several useful geometries per pair and adds independent restart/poll evidence before numerical stopping. Automatic foil screening admits a few verified reserve foils. Failed attempts do not count as usable pair coverage.

• Added optional normalized Cm preferences and a tailed-main-wing role. Flying wings retain local pitch-slope/static-margin checks; physical elevon trim is manual. Full diagnostic sweeps reach +20° with a separate scored alpha cap.
• Added explicitly user-started finalist sensitivity checks: fresh repeats, a third mesh, declared wake/transition/CG cases and additional speed points. Failed checks remain visible and original search scores remain intact.
• New candidate and export display names use wing_000123. Legacy source IDs, journals, checkpoints and native project bytes remain traceable and unchanged.
• Excel retains comparison/geometry tables and adds moment components, pitch curves and separate sensitivity data. The previous large-export memory fix remains included.
• Added hover help to input controls and Wing Optimizer 0.3.0 to the GUI title/header. Consolidated obsolete installed QA/build files and the previous source snapshot into one recoverable update archive.

Verification: 357 tests passed, four GUI tabs checked at two window sizes, and a small fictional workbook verified with ten charts and no unexpected formula errors. No real solver work was started for this update. See UPDATE_0_3_GUIDE.md and VALIDATION.md. Updated solver round trips and physical accuracy require user-started validation; this update does not establish a global optimum or mesh convergence.

Historical v0.2.1 release

This release adds direct Flow5 geometry exploration after rough airfoil-library screening. The product/package is Wing Optimizer / wing-optimizer; the compatible command/module remains wingopt. Original DATs, the real section cache and personal settings are retained.

Changes

• Made Flow5 geometry run duration explicit and adjustable in hours for each start/resume, with 0 removing the runtime cap. Two hours remains the default; CLI --time-limit-hours complements the existing seconds option. Resume keeps saved aircraft/solver settings while accepting revised time/candidate caps.
• Interleaved early local geometry refinement with global foil-pair exploration, so promising feasible wings can be refined before the entire foil-pair pool has been sampled. Continued exploration and existing numerical stopping checks remain; longer runs do not prove a global optimum.
• Added rotating alternative-foil checks on promising refined geometry and CG after finer-mesh evidence exists. They continue after initial pair coverage, allowing improved feasible foil pairs to guide later geometry refinement while global exploration and common finer-mesh checks continue.
• Require continued alternative-pair trial attempts before numerical stationarity when the configured cycle has room for these checks. Attempt counts can include repeats or failures and do not establish unique validated foil coverage. Finer-mesh rechecks now draw from current-cycle coarse leaders so an older optimistic coarse score cannot prevent new wings from receiving a fine check.
• Organized new GUI results under results/runs/YYYY-MM-DD_HH-MM-SS, keeping reports, inputs, checkpoints and solver jobs with each run. Export buttons save Results.xlsx and Flow5/ plus its ZIP in the active run folder; repeated exports retain prior files with numbered names. Existing local results use the same layout with compatibility links at their previous paths. Original DATs, reusable section polars and manual solver controls are retained.
• Added a visible Wing geometry Excel sheet with saved dimensions, foil selections, CG and a sortable table for all candidates.
• Added Export top Flow5 projects… and automatic saved-3D project packaging alongside Excel. Run-local folders/ZIPs preserve original native projects, XML/DAT inputs and polar data for manual Flow5 work, without launching solvers.
• Fixed Excel export memory growth in long geometry runs. Chart formulas share one compact source-row lookup per selected point instead of repeatedly scanning full tables. All candidates, full saved polars and evidence remain available; default curves select ranked wings with the chosen sweep.
• Added Start Flow5 geometry search and Resume Flow5 geometry search…. Library screening supplies rough foil pools; fresh Flow5 whole-wing sweeps provide the geometry/CG score at saved slow, cruise and fast speeds.
• Scores use weighted L/D at required lifting CL, fixed-control pitch-slope and static-margin checks. No preliminary assumed-trim penalty enters this score. Results remain untrimmed and separate from final checked physical-trim ranking.
• Defaulted this workflow to VLM2 after an independent diagnostic found severe TRIUNIFORM refinement sensitivity. The benchmark qualifies this choice; it does not certify total drag or swept flying-wing accuracy.
• Added three independent starts by default, rotating global sampling, foil-pair coverage, continued local refinement and common finer-mesh rechecks. Numerical stationarity requires configured mesh agreement, comparable pair ordering and adequate exploration. Bound-hitting variables generate warnings.
• Added a default two-hour runtime cap per explicit start/resume and optional per-invocation candidate limits. Manual pause checkpoints after the active solver case. Budget limits, coarse-only results and mesh-sensitive candidates do not become convergence or final aircraft recommendations.
• Added CLI --time-limit-seconds and --evaluation-budget overrides, including on resume. New requested whole-wing evaluations always execute fresh; section polars remain reusable and checkpoints retain search history.
• Implemented version-bound 7.54 native foil-only archives with actual hinges and a strict saved-project applied-state reader. Real 0°/±4° pilots confirmed geometry/control transfer without modifying source DATs or creating fake aerodynamic rows. Physical trim and nonlinear coupling remain gated.
• Fixed Flow5 7.57's aligned header/first-row join in the polar parser. Recovery requires an exact finite numeric field boundary; malformed rows are rejected and lift CL remains separate from rolling moment Cl.
• Bundled portable reconstructed/digitized benchmark fixtures and saved solver metrics with process failures, original hashes and uncertainty labels.

Real validation evidence

User-authorized installed-solver work now includes a diverse VLM2 comparison: 30 jobs attempted, 28 completed with usable outputs. Two slow-speed cases for wing_01988 could not reach the required CL inside the solved lifting branch; four wings have all six slow/cruise/fast × mesh comparison conditions.

A separate geometry smoke run completed one candidate evaluation across three speeds, then stopped as evaluation_limited. It is coarse-only and supplies no final recommendation. Three native control pilots passed geometry/hinge/ analysis-settings readback; missing polar points remain failures.

The independent AR=3 benchmark found much smaller VLM2 mesh sensitivity than TRIUNIFORM. Saved 7.57 TRIUNIFORM coefficients matched 7.54 at exported precision for each mesh, while the 7.57 processes timed out and retain failed status. The viscous benchmark stopped after section convergence failures, leaving total drag accuracy unavailable. See FLOW5_ACCURACY_BENCHMARK.md and FLOW5_NATIVE_CONTROLS.md.

Final software verification passed 276 tests in 80.516 seconds, including 17 Python Excel tests. Real geometry export produced seven charts without unexpected export errors and retained the solver score. The checked synthetic workbook export regression produced 14 charts with ten selectors, plus geometry selection, zero/missing-value and source-reordering checks. Saved project export preserved four refined candidates and 24 native projects in a 5.39 MB ZIP. Native reopening in Flow5 was not performed. Native Excel geometry checks were skipped to preserve the user's open workbook; its 16 historical checks remain separate. Mocked solver tests and saved fixtures establish software behavior, not aerodynamic correctness. The full-suite log is validation/full_tests_geometry_handoff_20261008.txt in the workspace validation directory.

Current limits

No verified physical-trim provider is supplied. Checked production search and physical-trim repeatability remain gated until actual lift/moment trim, reserve, fixed-control derivatives, nonlinear coupling and complete required settings evidence are available. The native hinge pilot satisfies only part of that contract. A later solver version alone does not certify these capabilities.

Native Flow5 7.57 MOPSO completed a manual experimental GUI pilot: four particles, two iterations, VLM2, 14 m/s, fixed alpha 4°, Ncrit 5 and span 0.7–0.8 m as the sole optimization variable on the MH32/AG13 seed wing_01863. Runtime was 5.46 s; the displayed best was span 0.783 m, CL 0.322968, CD 0.0160862. The CL 0.34 target with zero tolerance was not reached. Store best created an Optimized plane. Its stored plane summary showed zero mass and zero CG, so aircraft mass/CG must be explicitly restored and checked before reuse. This proves a limited GUI route, not app/headless integration, physical trim, multi-speed convergence, accuracy or aircraft optimality.

Two meshes, repeated runs and numerical stationarity do not prove convergence, physical accuracy or a global optimum. Stall/separation, propwash, structure and flight behavior remain outside the verified scope. Propulsion/energy-dependent estimates need appropriate supplied data and verified calculations.

Historical v0.2.0 verification

The earlier build passed 221 unit tests in 73.425 seconds and 16 native Excel comparison checks. Its four-tab interface was checked at two window sizes with 53 settings controls. Those are historical software/export checks, not the current v0.2.1 full-suite count or real physical-trim validation.

The real Ncrit 5 library has 34,132 usable cached cases of 39,925 requested, 5,793 failures and zero pending. Historical read-only screening found 1,131 foils covering the test-mission Reynolds envelope. Selection depends on the current aircraft requirements and cache conditions.