Known issues in the published data files

The cleaned tables served by the API apply the corrections (units, signs, frames); the raw files in the repository do not. Do not apply the corrections again to the published files.

Contents
  1. Known issues in the published data files
    1. Across cases
    2. Case A
    3. Case B
    4. Case C
    5. Case D
    6. Case E
    7. Case F
    8. Case G
    9. Case H
    10. Case I
    11. Case J
    12. Case K
    13. Case L
    14. Case M

Across cases

Scalar vs vector mean speed

Cases C, D and E, and the tunnel part of Case F, measure with thermistor anemometers: an omnidirectional scalar speed, not |mean velocity vector|. The data paper suggests comparing \(\sqrt{\lvert\mathbf{U}\rvert^2 + 2k}\) as a conservative correction. Case D, over the 33 points both instruments share: thermistor / \(\lvert\mathbf{U}\rvert\) = 1.34, as a ratio of means. Compare the bracket \(\lvert\mathbf{U}\rvert\) ≤ \(V_{\mathrm{s}}\) ≤ \(\sqrt{\lvert\mathbf{U}\rvert^2 + 2k}\).

Each case has its own frame and unit

A–D and H–L: model-scale metres, \(+x\) along the wind at \(\mathrm{WD} = 0\). E, F, G and M: full-scale metres. E and F: \(+x\) east, \(+y\) north, meteorological (from-) directions. G: \(+x\) downwind. M: \(+x\) north, \(+y\) west. In F and M the heights are above the tunnel floor, not the ground. frame in /api/cases/{id}.json states it.

Concentrations are normalised per case

Cases H to M carry a tracer concentration C in ppm and concentration_ratio = \(C/c_0\). \(c_0\) is defined by each case’s own release rate, reference speed and length (normalisation in /api/cases/{id}.json). Case K: the workbook and its data paper use different reference scales (a factor of 23.6 apart). Case M: field values are published in ppq and converted to ppm. Do not compare concentration_ratio across cases without reading each case’s normalisation.

Case A

  • \(V\) is positive toward \(+y\) (\(V\) = −2.893 m/s at (−0.040, −0.050, 0.010), just outside the \(y \lt 0\) side face).
  • Coordinates are metres; turbulence columns are rms values. The API serves variances (\(\overline{u'u'}\) = \(u_{\mathrm{rms}}^2\)) so A and B share one column meaning.

Case B

  • \(V\) is positive toward \(-y\), the opposite of Case A. The API flips it (lateral_sign_applied: -1).
  • Coordinates are millimetres (x[mm]); velocities are m/s. Read as metres, every probe lands 1000× too far out. The API serves metres (coordinate_scale_applied: 0.001).
  • Geometry was taken from each case’s own measurement-point drawing (MP-z_case*.png). Both buildings are centred on the origin; points are published on the \(y \le 0\) half by the tunnel’s symmetry convention.
Data notes: Cases A and B, sign of V and geometry source

Just outside the windward side face the flow must deflect away from the centreline, toward \(-y\) on the \(y \lt 0\) side face.

  • Case A: \(V\) = −2.893 m/s at (−0.040, −0.050, 0.010), edge at \(y = -0.04\).
  • Case B: \(V\) = +3.630 m/s at (−0.025, −0.110, 0.013), edge at \(y = -0.100\).

Same flow feature, opposite sign, so Case B’s \(V\) is positive toward \(-y\). Scored without the flip, Case B’s \(V\) gives \(R = -0.910\) against an \(R = +0.959\) streamwise control on the same points, an anti-correlation no turbulence closure produces.

The geometry was not inferred from the CSV extents.

Case C

  • The CB column is text (0D, 1D, 2D); a reader that parses every field as a number gets NaN for it. The header ends with a trailing comma (an empty 8th column).
  • 1080 rows = 3 wind directions × 3 centre buildings × 120 points: nine configurations. In the 0D configuration the 25 centre grid nodes are still absent although the ground there is free.
  • The CSV and Figure 3 of the data paper disagree on the sense of the WD rotation. This release keeps the CSV frame: WD rotates the wind from \(+x\) toward \(-y\). \(\mathrm{WD} = 0\) is not affected.
  • 24 of the 120 points sit exactly on a building face plane (\(\lvert x\rvert = 0.30\) or \(\lvert y\rvert = 0.30\)), in the gap between blocks.
  • The 0.40 m pitch is read from the drawing and the paper (street = D); the point grid alone bounds it to (0.35, 0.40].
  • Repeatability floor from the mirror-symmetric pairs: mean \(\lvert\Delta V_{\mathrm{s}}\rvert\) 0.088–0.137 m/s (0.024–0.038 \(U_H\)).
Data notes: Sense of the WD rotation

In the CSV’s own frame the \(\mathrm{WD} = 45^\circ\) field is mirror-symmetric about \(y = -x\) to within the experiment’s repeatability (0.10–0.13 m/s) and not about \(y = +x\) (0.5–0.8 m/s), so WD rotates the wind from \(+x\) toward \(-y\). Figure 3 of the data paper draws the fan the other way. The question is invisible at \(\mathrm{WD} = 0\).

The solved field settles it: scored in the CSV’s sense, R against the tunnel is 0.83–0.91 at 22.5° and 45°; in the figure’s sense it falls to −0.33 … +0.03.

Case D

  • All the data lives in LF_caseD.xls, a BIFF8 workbook (xlrd 2.x reads it; openpyxl cannot). Coordinates are millimetres.
  • The shipped BD_caseD.stl has 1572 facets, 576 of them duplicates, and all 83 roofs wound inward (signed volume 0.011159 m³ against the true 0.021477 m³). The API’s geometry.stl is rebuilt from the 83 shells’ boxes: 996 facets, watertight, volume 0.021477 m³.
  • Unmeasured thermistor cells are - (null in the API). At \(\mathrm{WD} = 0\) the 66 measured values are 33 exact mirror pairs.
  • The data do not fix the rotation sense for 22.5° and 45° (the array is exactly y-mirror symmetric, so \(+\theta\) and \(-\theta\) give exact mirror fields). This release rotates \(+x\) toward \(+y\), following Figure 4 of the data paper. Solved Eddy3D fields score R 0.80 (22.5°) and 0.85 (45°) in that sense, 0.40 and 0.17 mirrored.
  • All 78 points lie outside every block and inside the array; shifting the cloud by 12.5 mm (5 m full scale) in x or y puts two points inside a block.
  • The Zenodo record (10.5281/zenodo.15402668) also ships CSVs: unmeasured is −999.9 there, and RS-S_caseD.csv’s z header is the full-width character z. They match the workbook to \(4 \times 10^{-16}\); the repository mirrors them under clean/sources/zenodo_caseD/.
  • case-d-eddy3d-prior-comparison.xlsx (kept in the folder) scores \(R = +0.096\) against the thermistor; the cause is a frame or geometry error. Do not use it as a validation.

Case E

  • The velocity ratio is normalised by the inflow velocity at 15.9 m (the Niigata meteorological station height): \(U(15.9\text{ m})\) = 0.504531 × \(U_R\) = 3.935 m/s with \(U_R\) = 7.8 m/s at \(Z_R\) = 250 m. The reference is neither \(U_H\) nor \(U_R\).
  • case-e-eddy3d-prior-comparison.xlsx (kept in the folder) divided by the wrong reference and is low by a factor of 2.048.
  • All 2560 published ratios are integer multiples of 0.0198204 (55 distinct levels): the data’s own resolution is ±0.0099, 3.1 % of the median and about 10 % at the lowest points. A tolerance below one quantum (0.0198204) is below the data’s resolution.
  • The readme describes AF_caseE.csv, RS_caseE.csv and BD_caseE.stl; the case folder holds only the .3dm, .ghx, .xls and the MP drawing. The Zenodo files are mirrored under clean/sources/zenodo_caseE/: RS_caseE.csv is the workbook rounded to two decimals, and BD_caseE.stl is model scale (×250), not watertight (5,253 open edges), with point 50 at (−45.5, −45.0) 0.17 m inside one of its walls (2.06 m clear in the 3dm). Everything served here is extracted from the workbook and the 3dm.
  • case-e.3dm is full-scale metres; MP_caseE.png is drawn at model scale (±0.8 m). Use layers buildings and buildings_new_construction; the map plate on the Japanese-named layer, the terrain plates and probes are not buildings. 1,563 of the 8,436 triangles (18.5 %) of its building meshes are wound against their neighbours; all are re-wound in the API’s STLs.
  • Do not take measurement coordinates from the probes layer: its spheres sit a median 2.31 m (max 21 m) from the workbook’s points, which are authoritative.
Data notes: Case E geometry
  • 546 of the 548 bodies of the after-construction mesh (543 of 545 before construction) have triangles wound against their neighbours; a body is a set of triangles joined through edges that exactly two triangles use.
  • 475 of the 503 building shells (478 of 506 after construction) are wound inward in the 3dm. A viewer that culls back faces draws the reversed triangles of a shell wound partly inward as missing walls.
  • Every building is now wound consistently (each edge is traversed once in each direction) with a positive volume: 545 closed bodies after construction, 542 before, the smallest 4.125 m³.
  • One wall that two buildings share (4.27 m × 3 m, x 3.5 to 7.5 m, y −76.5 to −75 m) is modelled twice, once per building, with different diagonals, so the copies never joined their buildings. The 4 triangles of both copies are removed and the two buildings are one body.
  • 11 edges where two separate buildings touch (15 before the wall is removed) are counted as body_contacts; each building still sees them as ordinary two-triangle edges, so a contact does not open a body.
  • geometry.json reports bodies, edges_boundary, body_contacts, inconsistent_triangles and min_body_volume; watertight is true when every body is closed, consistently wound and positive in volume.

Case F

  • The 16 field rows of point 14 in RS-FO_caseF.csv hold the workbook's value of the NEXT wind direction clockwise (the series is shifted by one); the cleaned table follows the workbook (clean/QUESTIONS.md QF1).
  • Tunnel and STL coordinates are MODEL scale (1:800); the field's are full scale. The case frame is full scale; the tunnel's \(x\), \(y\), \(z\) and the STL are multiplied by 800.
  • z is height above the tunnel floor. The readme's rule 'ground = z - 10 m' gives the 5 / 9 / 20 m planes, but the STL's terrain under all 13 points is at \(z = 5\thinspace\mathrm{m}\), so the planes stand 10 / 14 / 25 m above it.
  • BD_caseF.stl is wound inward on every ground, roof and sampled wall facet and is open along the rim of its 725 m disc (812 boundary edges); the served geometry.stl is re-wound outward and closed with a base slab down to \(z = -1\thinspace\mathrm{m}\).
  • MP_caseF.csv names the two reference points the other way round from the workbook, the STL and MP_caseF.png (the workbook's Shinjuku-Mitsui Building is the 225 m tower at (-17.5, 96.0), the KDD Building the 180 m one at (85.8, -347.1)).
  • Field values are the legacy workbook's roof-anemometer ratios times 0.940 (NW-NE) or 0.963 (other directions); sigma is the workbook's S.D. rounded to 0.01 times that factor, so it moves in steps of 0.0094 / 0.0096.
  • The tunnel did not run E, ESE, WSW or W; the field data cover all 16. The 5 m and 20 m planes exist at some points only (29 and 79 rows).
  • One tunnel value (point 8, SW, 9 m: 0.474) has no counterpart in the workbook's results sheet, and the flat 'Sheet1' table has 0.63 there.
  • Field anemometer heights above the tunnel floor are 13 - 21 m (3 - 11 m above ground by the readme's rule); the data paper says 3 - 9 m, so points 13 and 14 (11 m) disagree.
  • The thermistor reads a SCALAR speed and the cup anemometer a 10-minute mean: a RANS \(\lvert\mathbf{U}\rvert\) is biased low against both.
  • The files do not say whether a wind direction names where the wind comes from. The cleaned tables read it as the direction the wind comes from (N = 0, E = 90). At NNW a solved field supports that reading: Eddy3D's OpenFOAM 12 run for the wind from NNW correlates R 0.935 with the tunnel's 9 m column labelled NNW and R 0.077 with its column labelled SSE, the opposite wind (clean/QUESTIONS.md QF9); the other directions are not solved and stay unverified.

Case G

  • FIELD data: full-scale metres, no model, and nothing in m/s. The table carries \(U/U_H\) (velocity_ratio) and \(k/U_H^2\) (k_ratio), \(U_H\) = the mean speed at 7 m during each period; the approach-flow table carries \(U/U_b\) and \(k/U_b^2\) (\(U_b\) = \(U\) at 9 m) in its \(U\) and \(k\) columns, which are dimensionless here.
  • The readme lists the heights of the lee measurement as 1.5, 3.0, 6.0, 9.0 m; the table, the workbook, the drawing and the data paper have 1.5, 3, 4.5, 6 m.
  • The tree belt is a porous canopy (shading rate 81 % after pruning): the served geometry is its 2 m × 74 m × 7 m envelope with the 1.2 m gap, and a solver needs a canopy (drag) model.
  • The approach profile (28 Dec 2000, 120 min) and the lee measurements (26 Dec 2000, 10 min or more) are from different days. Behind the belt \(k/U_H^2\) is 0.040-0.072 at \(5\thinspace H\), 3-8 times below the approach flow's 0.26-0.31 at the same heights; the files do not say why.
  • Only 28 numbers per quantity exist, on one centre section, one time-average each, with no repeats and no uncertainty.
  • The wind direction relative to the belt normal during the measurement is not published; the case is two-dimensional by assumption.

Case H

  • The table carries the legacy workbook's values (ratios at full precision x \(u_H\) = 3.226 m/s), not RS_caseH.csv's: on the \(z/H = 1/2\text{ and }1/4\) planes (and in the approach profile) the CSV's \(U\), \(V\), \(W\) are the workbook ratios ROUNDED TO 0.01 before the scaling, a 0.032 m/s grid (+/-0.016 m/s).
  • One row per instrument and point: velocity rows (split fibre) and concentration rows (hydrocarbon analyser) with the same point_id (the CSV's No.). The velocity line is \(x = 0.40\thinspace\mathrm{m}\) and the concentration line \(x = 0.38\thinspace\mathrm{m}\), so 50 points have no velocity and 50 no concentration (-999.9 in the CSV); none of them is inside the building.
  • \(V\) of the vertical-plane sheet (point numbers 1-220) is the exact negative of \(V\) at the same points in the horizontal-plane sheets (positive toward -y); it is multiplied by -1 so that \(V\) is positive toward \(+y\) everywhere, as in Cases A and D.
  • The \(y = 0\), \(z = 0.1\) line (13 locations) is in two sheets and appears twice in the table, as exact copies once \(V\) is flipped; they are not independent repeat measurements.
  • Concentration is the analyser's reading: upwind of the building (where the true value is zero) it is +/-0.1 ppm, and 97 of 826 values are negative (minimum -0.83 ppm). Log-based statistics (geometric mean, geometric variance) need a floor.
  • The published header misspells \(w_{\mathrm{rms}}\) as 'w_rsm'; the workbook's inflow sheet labels z as 'mm' but holds metres.
  • The profile's \(\overline{u'w'}\) and \(\overline{u'v'}\) are published but the approach-flow table has no column for them.

Case I

  • \(V\) on the \(y = 0\) plane (53 of 91 points) is 0 in the CSV and 1e-9 … 4e-6 m/s in the workbook, a symmetry value. It is NaN in the table; \(V\) is measured only at the 38 off-plane points.
  • The table carries the legacy workbook's values; the CSVs round to 3 decimals (velocity, \(k\)) and 1 decimal (concentration, ppm). Every concentration is a multiple of 0.1 ppm: a 0.4 ppm value carries +/-12 % of quantisation.
  • The paper's \(c_0\) = 39.2 ppm needs the ETHANE flow (10 % of the 400 cc/min, 6.67e-7 m³/s); its sentence 'ethane released at 6.67e-6 m³/s' is the mixture. With the mixture flow \(c_0\) would be 392 ppm.
  • The workbook normalises the approach flow with 1.697 m/s and the points with 1.7 m/s (0.18 % apart).
  • AF_caseI.csv ends in 53 rows of ',,' and two stray cells, '0.5*' and '1.5*' (chart axis labels): a reader that does not test z for a number takes them for data.
  • Thirteen ground-centreline locations (x, 0, 0) are measured twice (they are on both the \(y = 0\) and the \(z = 0\) plane) with values up to 8.5 % apart; both are kept. The four \(y = 0\), \(z = 0.05\) wind points that appear twice in the workbook are identical copies and appear once.
  • Only the TKE is published, no variances and no concentration fluctuation: \(k\) cannot be checked against the components.

Case J

  • Temperature is in KELVIN in all three published CSVs; the API serves degrees Celsius (T) and kelvin differences (\(T_{\mathrm{rms}}\), heat fluxes).
  • The published CSV is the legacy workbook's non-dimensional values rounded to 3-4 decimals and scaled back; the table is rebuilt from the workbook. The CSV's heat fluxes sit on a 0.0049 K m/s grid and \(k\) on 0.00021 m²/s².
  • \(c_0\) = 12.283 ppm includes the 5 % ethylene fraction of the released gas (workbook cell Co); \(q/(U_H H^2)\) alone is 245.7 ppm. The data paper states no \(c_0\) for Case J.
  • RS_caseJ.csv lists the location (-0.16, 0, 0.004) twice (No. 1 and 153, identical values): a score over the table counts that point twice.
  • The workbook's \(\overline{u'^2}\) is not always the square of its \(\sigma_u\) (up to 3 % in single rows); the table uses the variances, which reproduce \(k\).
  • The workbook lists an expanded uncertainty beside every quantity (no column here; clean/manifest.json cases.J.extra_measurements.expanded_uncertainty). The mean concentration is below its own uncertainty (about 1 ppm) at 42 % of the rows, and 193 uncertainty cells of the w'theta' flux are an Excel fill series, not values.
  • The empty-tunnel profile changes along the tunnel: \(U\) at \(z = 0.019\thinspace H\) is 0.33 \(U_H\) at \(x/H = -2.5\) (the published AF_caseJ.csv), 0.58 at the building's own position. The workbook has all three stations (clean/cases/J/approach_flow_full.csv).
  • The \(+y\) and -y halves are not mirror images: over the 97 mirror pairs of the \(z/H = 0.25\) plane \(U\) at \(+y\) is 11 % lower (means 0.558 and 0.629 \(U_H\), R between the halves 0.955), so a y-symmetric model has a ceiling on this file.
  • The readme's Reynolds number (43,000) is Case H's; \(U_H H/\nu\) = 1.6e4 as the data paper states.
  • \(\overline{u'v'}\), \(\overline{u'w'}\), \(\overline{v'w'}\) are not published at the points; \(\overline{u'c'}\), \(\overline{v'c'}\), \(\overline{w'c'}\) were not measured.

Case K

  • The legacy workbook normalises with \(H = 0.2\thinspace\mathrm{m}\) (Case H's building), \(U_H\) = 4.94009 m/s and \(C_0\) = 18.218 ppm, not the cube's \(H = 0.06\thinspace\mathrm{m}\) and \(U_H\) = 2.33 m/s; its normalised columns are 23.6 times smaller in concentration. The API's concentration_ratio uses the cube's \(c_0\) = 429.18 ppm.
  • RS_caseK.csv has no point numbers; point_id is the row order of the file.
  • The published CSV is the workbook rounded (3 decimals, concentrations 2), with \(k\) recomputed from the rounded rms values; the table is rebuilt from the workbook.
  • AF_caseK.csv is Case H's approach-flow file, row for row (\(x = -0.493\thinspace\mathrm{m}\), tabulated for a 0.2 m building); it gives 2.24 m/s at \(z = 0.06\thinspace\mathrm{m}\) against the stated \(U_H\) = 2.33.
  • The lateral velocity \(V\) is dominated by a cross-flow offset (about 0.06 m/s, also in the empty tunnel); the \(+y\) and -y halves cannot confirm its sign.
  • The mean concentration is not mirror-symmetric (the \(+y\) and -y values of one point differ by 44 % on average), so the mirror pairs are not a repeatability estimate; velocity and \(k\) are (3 % and 5 %).
  • The workbook's \(\langle u_{\mathrm{sc}}\rangle\) / \(\sigma_{\mathrm{usc}}\) ('velocity magnitude', never below \(\lvert\mathbf{U}\rvert\)) is not in the table.
  • Data paper section 3.11 describes the measurements as 'approximately 30 points at eight heights downwind of the line source'; Case K has 203 points at five heights and a point source.
  • Workbook units: x and y in millimetres in both data sheets; z in millimetres in 'Measurement data' but metres in 'Inflow'.

Case L

  • RS_caseL.csv's temperature column is headed 'T(K)' but holds DEGREES CELSIUS (12.85 .. 29.86); AF_caseL.csv's 'T(K)' is kelvin (283 .. 295). The API serves degrees Celsius for both.
  • The published CSV is the workbook rounded to 2-4 decimals; the table is rebuilt from the workbook. Its 'tc' column (temperature-concentration covariance, K ppm) and the workbook's 'Usc' (the vector magnitude of the mean velocity) are not in the table.
  • BD_caseL.stl: the two bottom triangles (\(z = 0\)) of each of the 126 cubes (252 triangles) are wound inward; the served geometry.stl is the 126 boxes, outward.
  • The readme's file list names AF_caseK.csv and RS_caseK.csv; the files are AF_caseL.csv and RS_caseL.csv.
  • The readme's 'building surface temperature \(T_b\) = 14.8 degC' is the approach-flow air temperature at \(z = H\); the workbook gives the building surface as 42.2 degC.
  • The line source is 0.037 m x 1.2 m in the readme and data paper but a 0.04 m x 1.02 m patch in BD_caseL.stl.
  • The \(+y\) and -y halves are not mirror images: the same point differs in \(U\) by 26 %, in \(k\) by 23 %, in C by 36 % on average, and \(U\) is 0.11 m/s lower and T 1.3 K higher at \(+y\) at \(z = 0.15\thinspace\mathrm{m}\) (0.09 m above the roofs), so the tunnel flow has a lateral gradient; do not assume symmetry.
  • The workbook carries no expanded uncertainty (the data paper mentions one for Case J only).
  • AF_caseL.csv lists the profile from the top (\(z = 0.32\thinspace\mathrm{m}\)) down and has no \(k\).
  • The measured tracer flux through the cross-sections equals q / 1.2 m to 0.86 .. 1.0 (street and cube-column lines averaged), which fixes q and the unit ppm; the cube-column line alone carries only 0.2 .. 0.4 of it.

Case M

  • Field concentrations are in ppq (1e-15 volume fraction) and the tunnel's in ppm: 1 ppq = 1e-9 ppm. The cleaned table converts the field's to ppm and carries \(C/c_0\) for both, with the run's own \(c_0\) in the field.
  • Every field run has its own reference speed \(U_R\), mean direction \(\mathrm{WD_R}\) and tracer flow q (clean/manifest.json -> cases.M.extra_measurements.field_runs); the row's wind_direction_deg is \(\mathrm{WD_R}\).
  • The tunnel measured velocity only at 180°; at 175° and 185° the CSV has -999.9 in all seven velocity columns.
  • RS-WT_caseM.csv is the legacy workbook rounded (2 dp velocity, 1 dp concentration), and its \(k\) is half the sum of the 2-dp RMS values, up to 0.005 m2/s2 from the workbook's; the cleaned table carries the workbook's values.
  • The tunnel source is on the ground in the workbook, LC_caseM.csv and the STL (readme: 'on a rooftop'); the field source was 1.5 m above ground.
  • \(H_R\) = 41 m is the anemometer height above the ground; the main roof under the reference point is about 33 m above it. \(c_0\) uses \(H_R\) = 0.068 m (41 m / 600 = 0.06833 m would give 131.2 ppm, -1 %).
  • The field has an ambient floor of 8 - 11 ppq at points 11 - 15 in every run (not subtracted); the tunnel's concentration has a zero offset (two negative means at point 12) and an rms floor of 7 ppm.
  • Only 11 of the 18 field runs are within 5° of 180°, the tunnel's direction; runs 1, 2 and 9 have \(U_R\) of 1.4 - 1.8 m/s.
  • BD_caseM.stl is MODEL scale, wound inward on its ground and roof facets, and has 21 boundary and 7 non-manifold edges (7 on the rim of the panel, 14 on 7 slivers hanging on building roofs); the served geometry.stl is multiplied by 600, re-wound outward, without the slivers and closed with a base slab down to \(z = -1\thinspace\mathrm{m}\).
  • The approach-flow heights are above the empty tunnel floor: at the profile station the model's taper is 0.052 m high, above the 10 lowest profile heights.