Document native-point surface quadrature sidecars

Adds the file contract, area-weighted metric example, selective download command, and release note for all 1,800 dual-area sidecars.

README.md

@@ -35,6 +35,7 @@ In order to cite the use of this dataset, please cite the corresponding preprint
 Each folder (e.g., `geo_LHCi_AoA_j` where `i` is the geometry ID from 001 to 180 and `j` is the angle of attack) corresponds to a different geometry and angle of attack. Inside each run folder, you will find:
 
 * `boundary_geo_LHCi_AoA_j.vtu.tgz`: Time-averaged flow quantities on the surface boundary (approx. 13 GB compressed, 21 GB unzipped).
+* `boundary_dual_area_geo_LHCi_AoA_j.npy`: Native-point surface quadrature weights for the corresponding boundary VTU (little-endian float32, one value per VTU point, in exact VTU point order, with units of `in^2`).
 * `volume_geo_LHCi_AoA_j.vtu.tgz`: Time-averaged flow quantities within the domain volume (approx. 23 GB compressed, 86 GB unzipped).
 * `geo_LHCi_AoA_j.stl`: Surface mesh (tris) of the geometry (approx. 197 MB).
 * `geo_LHCi_AoA_j.stp`: Surface geometry definition (approx. 48 MB).
@@ -50,6 +51,7 @@ In addition to the files per run folder, there are also general summary files in
 * `geo_values_all.csv`: Reference geometry values used to define the geometry via the DoE method for all runs.
 * `force_mom_all.csv`: Time-averaged drag, lift, moment, and pressure/viscous coefficients for all runs.
 * [`splits/`](splits/): Deterministic train/validation/test split manifest, documentation, and generation code.
+* [`surface_dual_areas/`](surface_dual_areas/): Definition, generator, complete 1,800-case manifest, per-case topology/provenance records, validation summaries, and SHA-256 hashes for the native-point surface quadrature sidecars.
 
 ## Recommended Dataset Splits
 
@@ -64,9 +66,41 @@ For reproducible machine-learning evaluation, HiLiftAeroML provides 14 determini
 
 The levels within each data-efficiency ladder are nested and share the same validation and test sets, enabling direct comparisons across training-set sizes. Use validation data for model selection and reserve test data for final evaluation. See [`splits/README.md`](splits/README.md) for the complete definitions, counts, and selection guidance, or [`splits/generate_splits.py`](splits/generate_splits.py) for the generation and validation logic.
 
+## Area-Weighted Evaluation on the Native Surface
+
+Every raw boundary VTU now has a matching `boundary_dual_area_<case>.npy` sidecar. If a native surface polygon is fan-triangulated, each triangle contributes one third of its area to each incident vertex. The resulting mass-lumped dual areas sum to the fan-triangulated surface area and allow point-data predictions to be scored by physical surface measure rather than by the nonuniform density of mesh points.
+
+The sidecar applies **only** to the corresponding raw `boundary_<case>.vtu`: its length equals `mesh.n_points`, and entry `i` weights point-data entry `i`. Do not use it after reordering, resampling, converting to STL, or processing the mesh into another representation. Accumulate statistics in float64 even though the stored weights are float32. For a scalar or vector field, an area-weighted relative L2 error can be computed as follows:
+
+```python
+import numpy as np
+import pyvista as pv
+
+case = "geo_LHC001_AoA_4"
+mesh = pv.read(f"{case}/boundary_{case}.vtu")
+w = np.load(
+    f"{case}/boundary_dual_area_{case}.npy", mmap_mode="r"
+)
+assert w.dtype == np.dtype("<f4") and w.shape == (mesh.n_points,)
+
+# pred and reference have shape (mesh.n_points,) for a scalar field or
+# (mesh.n_points, n_components) for a vector field.
+err2 = np.square(pred.astype(np.float64) - reference.astype(np.float64))
+ref2 = np.square(reference.astype(np.float64))
+if err2.ndim == 2:
+    err2 = err2.sum(axis=1)
+    ref2 = ref2.sum(axis=1)
+
+numerator = np.sum(w * err2, dtype=np.float64)
+denominator = np.sum(w * ref2, dtype=np.float64)
+relative_l2 = np.sqrt(numerator / denominator)
+```
+
+The weights are in square inches because the native coordinates are in inches. A constant unit conversion cancels from normalized metrics such as relative L2 and weighted means. See [`surface_dual_areas/README.md`](surface_dual_areas/README.md) and [`surface_dual_areas/manifest.json`](surface_dual_areas/manifest.json) for the exact algorithm and validation contract.
+
 ## How to Download
 
-HiLiftAeroML is publicly available on Hugging Face. There is no Hugging Face access charge, but the dataset is very large: the full repository is about 65.9 TB as stored on Hugging Face and substantially larger after extracting the `.tgz` archives. Please make sure you have sufficient storage and bandwidth before downloading large subsets.
+HiLiftAeroML is publicly available on Hugging Face. There is no Hugging Face access charge, but the dataset is very large: the full repository is about 66.9 TB as stored on Hugging Face and substantially larger after extracting the `.tgz` archives. Please make sure you have sufficient storage and bandwidth before downloading large subsets.
 
 We recommend using the Hugging Face CLI rather than `git clone` for selective downloads, resumable local downloads, dry-run previews, and Xet-backed large-file support.
 
@@ -122,7 +156,22 @@ hf download nvidia/HiLiftAeroML \
 
 Remove `--dry-run` to download those files.
 
-### Example 3: Download Specific Runs with a Bash Script
+### Example 3: Download Only the Surface Quadrature Sidecars
+
+The following command downloads the 1,800 native-point NPY files and their algorithm, provenance, and validation metadata without downloading the CFD archives. The sidecars occupy about 1.02 TB (946.7 GiB) in total, so preview the selection first.
+
+```bash
+hf download nvidia/HiLiftAeroML \
+  --repo-type dataset \
+  --local-dir ./hiliftaeroml_data \
+  --include "geo_LHC*_AoA_*/boundary_dual_area_*.npy" \
+  --include "surface_dual_areas/**" \
+  --dry-run
+```
+
+Remove `--dry-run` to download those files.
+
+### Example 4: Download Specific Runs with a Bash Script
 
 This script loops over geometry IDs 1-180 and AoA values 4-22 in steps of 2. By default it runs in preview mode. Set `DRY_RUN=0` to download the files.
 
@@ -185,6 +234,8 @@ This dataset is provided under the permissive open-source **CC-BY-4.0** license.
 
 ## Version History
 
+* 07/08/2026 - added native-boundary-point dual-area NPY sidecars for all 1,800 cases, together with a complete manifest, hashes, validation records, and generation code for reproducible area-weighted surface evaluation
+
 * 06/07/2026 - updated the splits to include a couple of extra ones (geometry_medium, scarce, super_scare and medium)
 
 * 26/05/2026 - IDs 0-43 for angles 16-22 had incomplete force_mom (didn't have std dev etc). Corrected. Please download forces again if you took files before this date