Flow5 accuracy benchmark

The current results do not establish full aerodynamic accuracy. On the tested rectangular wing, Flow5 7.54 VLM2 has much smaller mesh sensitivity than thin-surface TRIUNIFORM. TRIUNIFORM's lift slope collapses as the mesh is refined in both the saved 7.54 and 7.57 outputs. The viscous benchmark did not produce a usable whole-wing polar, so total drag accuracy remains untested.

Independent experimental reference

The reference is the UIUC rectangular AR=3 rounded flat-plate wing at Re=80,000 in Ananda, Sukumar and Selig (2015), Figure 6. The primary paper and DOI identify the experiment. The CSV is digitized from the figure's UIUC symbols, not raw author data; increasing and decreasing alpha branches remain separate.

Geometry: span 0.2667 m, constant chord 0.0889 m, area 0.02370963 m², no sweep, twist or dihedral. The 3.81 mm thick section has the published 5:1 leading-edge and 10:1 trailing-edge elliptical rounding. The 241-point DAT is an analytic reconstruction of that published geometry, not the original measured surface. Its sampled maximum polygon deviation from the analytic shape is about 13.4 µm; this is not a solver surface-accuracy proof.

The declared fluid properties are density 1.225 kg/m³ and kinematic viscosity 1.5×10⁻⁵ m²/s. Velocity is 13.49831271 m/s, computed from Re×ν/chord. Cm is referenced to quarter chord, x=0.022225 m. The arbitrary 1 kg mass and zero inertias do not impose an aircraft lift, trim or dynamic-stability constraint in these fixed-speed wing-only cases.

The paper reports absolute coefficient uncertainties CL=0.0103, CD=0.00107 and Cm=0.00024. Estimated digitization bounds are alpha=0.15°, CL=0.01, CD=0.0015 and Cm=0.0015. These plot-reading bounds are estimates, not statistical confidence intervals. The Cm axis is inverted in the figure: negative at the top, positive at the bottom. Post-stall CSV rows are retained for context and excluded from the comparisons below.

Saved solver results

All nine inviscid saved polars contain 22 distinct alpha rows from −4° to 6.5° at 0.5° increments. The table compares 17 rows from −4° to 4°, with linear interpolation onto each experimental branch and no extrapolation. RMSE ranges span the two branches. Every CD in these inviscid results is induced drag only; comparing it with experimental total CD would be misleading.

Mesh labels give requested semispan × chord panel counts. TRIUNIFORM has more elements than VLM2 for the same section panel settings, so this table does not claim an equal-cost comparison.

The experimental fitted lift slopes are about 0.0590–0.0593 per degree. VLM2's slope is lower and its quarter-chord Cm disagreement remains material; small mesh changes alone do not establish physical accuracy. Across the full −4° to 6.5° saved range, VLM2's maximum CL changes are 0.00289 then 0.00130, and maximum Cm changes are 0.0001022 then 0.0001692. These are numerical sensitivity measurements, not a mesh-convergence certificate.

The corresponding TRIUNIFORM maximum CL changes are 0.13632 then 0.198584, and maximum Cm changes are 0.0515305 then 0.07607. The parsed CL/CD/Cm rows in the 7.57 outputs match the 7.54 TRIUNIFORM rows exactly at the exported precision for each mesh. Upgrading alone did not resolve this benchmark behavior.

The TRIUNIFORM CSV's global “Dirichlet” label does not diagnose the cause. The official 7.54 P3UniAnalysis source explicitly uses Neumann for mid-surface panels via bNeumann() || isMidPanel(). The XML reader has no boundary-condition override field. Earlier scratch commentary attributing the collapse to that header label is superseded by this source check. The cause of the severe refinement sensitivity remains unresolved.

Failures and recovered artifacts

The 7.57 processes failed to exit within 180 seconds. Their original case statuses remain failed; their original benchmark_results.json files were not changed. Solver logs report that panel analysis completed, and saved whole-wing polars plus native projects are present. Those artifacts support the numerical comparison above, not a successful end-to-end process claim. Full native applied-state verification for these benchmark projects is absent.

Flow5 7.57 also joined the first numeric row onto its aligned CSV header. The parser now recovers it only when the header has N nonnumeric column labels followed by exactly N finite numeric fields. It preserves case-sensitive lift CL versus rolling moment Cl, rejects malformed boundaries and still checks negative drag, finite coefficients and duplicate alpha. Raw solver files were left untouched; exported fixture CSVs contain only their parsed numeric rows.

No auto-close XML field was found in the inspected official schema. The 7.57 command-line path runs executeScript synchronously and then marks the application done. MainFrame exports results and saves the project before returning. The saved logs lack the final “Script completed” marker. A post-analysis/export/exit-path issue is consistent with that evidence; its exact cause is not established.

The attempted 7.54 viscous sensitivity series stopped after systematic section failure. The first coarse case timed out after 300 seconds, with all 20 tested strips on its first surface failing XFoil on-the-fly at alpha=−4°, Re=80,000. The second case was interrupted and four remaining cases were not launched. There is no viscous whole-wing result for a total-drag accuracy conclusion. Ncrit was not measured in the experiment and is irrelevant to the inviscid results. Nonlinear viscous lift coupling was disabled throughout.

Portable evidence and practical interpretation

tests/fixtures/uiuc_ar3/ contains the reconstructed DAT, digitized reference CSV, original reference/uncertainty manifest, a small raw 7.57 parser fixture, nine parsed solver polars and saved_output_analysis.json. The analysis JSON retains raw polar, native project and log hashes, original process failures, coverage and recomputed bias/RMSE. summarize_saved_outputs.py recomputes it from saved research files only; it never launches Flow5. The paper PDF and large native projects are deliberately not bundled in the fixture directory.

Original research run IDs are 20261008T104825Z (7.54 TRIUNIFORM), 20261008T105617Z (7.54 VLM2), 20261008T105842Z (7.57 TRIUNIFORM timeouts) and 20261008T104042Z (stopped viscous series). They remain under the workspace's validation/benchmark_research/flow5_validation/ directory.

This supports using VLM2 as the current geometry-search method while retaining fresh runs, explicit mesh checks and failure records. It does not validate flying-wing trim, fixed-control derivatives, high-sweep behavior, separation, stall or a globally optimal aircraft. A verified real viscous and trimmed workflow still needs an appropriate independent benchmark.