{"success":true,"data":{"author":"Neil Ashton et al. (NVIDIA and collaborators)","bucketAccessType":"public","bucketName":"nvidia/HiLiftAeroML","bucketPrefix":"","category":"Physics","cloudProvider":"huggingface","createdAt":"2026-11-18T00:06:05.114Z","description":"High-fidelity computational fluid dynamics dataset of a high-lift transport aircraft, built for training and testing AI surrogate models in aerospace aerodynamics. It holds 1,800 cases: 180 geometry variants of the NASA Common Research Model high-lift configuration (CRM-HL), sampled by design of experiments, each run at 10 angles of attack from 4 to 22 degrees. Flows were computed with the Fidelity Charles explicit compressible finite-volume solver using wall-modeled large-eddy simulation (WMLES, dynamic Smagorinsky subgrid model, equilibrium wall model) on Voronoi meshes of 300 to 500 million cells. Each geo_LHC<i>_AoA_<j> folder holds time-averaged surface fields (boundary VTU, about 13 GB compressed) and volume fields (volume VTU, about 23 GB compressed), the STL and STEP geometry, force and moment coefficients (lift, drag, pitching moment with pressure and viscous parts), reference values, convergence plots and surface quadrature weights. Root files give all-run force tables, geometry parameters and 14 deterministic train/validation/test split families, including unseen-geometry, high-angle and post-stall extrapolation. The repository is about 67 TB; read the VTU files with PyVista or VTK and download subsets with the Hugging Face CLI. Released by NVIDIA and collaborators under CC-BY-4.0; cite Ashton et al. 2026, HiLiftAeroML (arXiv:2605.19565).","displayTitle":"HiLiftAeroML: Wall-Modeled LES of the NASA CRM High-Lift Aircraft for AI Surrogates","huggingFaceUrl":"https://huggingface.co/datasets/nvidia/HiLiftAeroML","id":"c330dcdc-bac4-5302-8388-6149389dc9a2","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/nvidia/HiLiftAeroML","objectCountEstimate":23423,"probeStatus":"ok","referenceUrl":"https://arxiv.org/abs/2605.19565","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"nvidia/HiLiftAeroML\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"nvidia/HiLiftAeroML\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=nvidia, license:cc-by-4.0, downloads=41993"],"format":"mixed","generatedAt":"2026-11-18T00:09:31.121Z","layout":{"example":"data_quality/volume_zero_fill_raw_point_ids_all1800.jsonl","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"{\"aoa_degrees\":4,\"case_id\":\"geo_LHC001_AoA_4\",\"case_index\":0,\"geometry_id\":1,\"raw_point_count\":224743190,\"zero_fill_point_count\":68,\"zero_fill_raw_ids_sha256\":\"3863ba027964438ac83a3fc8b0fecb61a65af577529594e4834ffbbdd42c3c6b\",\"zero_fill_raw_point_ids\":[10982626,10982627,10982628,10982629,10982630,10982631,10982632,47794847,47794849,47794851,47794853,47794855,47794857,47794859,73475956,73475959,73476061,73476064,73476313,73476316,73476414,73645724,73646089,73796011,73796785,74134204,74134211,74422804,1505233","encoding":"utf8","key":"data_quality/volume_zero_fill_raw_point_ids_all1800.jsonl","truncatedAt":512}],"sampleObjects":["data_quality/SHA256SUMS","data_quality/volume_zero_fill_raw_point_ids_all1800.jsonl","data_quality/volume_zero_fill_raw_point_ids_all1800_manifest.json","data_quality/volume_zero_fill_raw_point_xyz_all1800.csv","force_mom_all.csv"],"sidecars":[{"bytes":"---\nlicense: cc-by-4.0\nviewer: false\n---\n\n# HiLiftAeroML: High-Fidelity Computational Fluid Dynamics Dataset for High-Lift Aircraft Aerodynamics\n\n**Contact:**\nNeil Ashton (contact@caemldatasets.org)\n\n## Summary\nThis dataset provides the first-ever open-source high-fidelity CFD dataset of a high-lift aircraft for the purpose of AI surrogate model development. The dataset is composed of 1,800 samples, arising from 180 geometry variants of the NASA Common Research Model (CRM-HL) across 10 angles of attack (from 4° to 22° in 2° increments). \n\nOne of the key novelties of this dataset is the use of a GPU-accelerated high-fidelity explicit, wall-modeled LES (WMLES) approach on solution-adapted grids containing between 300M and 500M cells. By making this data publicly available, the aim is to accelerate the research and development of AI surrogate modeling within the aerospace industry by addressing the lack of high-fidelity, scale-resolving CFD training data for complex 3D airframes in the high-lift regime.\n\n## CFD Solver\nThe simulations were performed using the **\"Fidelity Charles\"** flow-solver, an explicit, unstructured, finite-volume solver for the compressible Navier-Stokes equations. \n* **Accuracy:** 2nd-order accurate in space and 3rd-order accurate in time. \n* **Meshing:** The domain was discretized using \"Fidelity Stitch\", an unstructured mesh generator based on Voronoi diagrams. \n* **Turbulence Modeling:** To capture complex, unsteady flow phenomena such as massive flow separation, a wall-modeled LES using the Dynamic Smagorinsky subgrid-scale model and an equilibrium wall model were employed.\n\n## How to Cite This Dataset\nIn order to cite the use of this dataset, please cite the corresponding preprint paper (https://arxiv.org/abs/2605.19565):\n\n```bibtex\n@article{ashton2026hiliftaeroml,\n  title={{HiLiftAeroML: High-Fidelity Computational Fluid Dynamics Dataset for High-Lift Aircraft Aerodynamics}},\n  year={2026},\n  journal = {arxiv.org}, url={https://arxiv.org/abs/2605.19565}, \n  author={Ashton, Neil and Clark, Adam and Heidt, Liam and Ivey, Christopher and Bose, Sanjeeb and Agrawal, Rahul and Goc, Konrad and Ranade, Rishi and Adams, Corey and Sharpe, Peter and Nidhan, Sheel and Akkurt, Semit and Leibovici, Daniel and Kossaifi, Jean}\n}\n```\n\n## Dataset Structure and Files\nEach 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:\n\n* `boundary_geo_LHCi_AoA_j.vtu.tgz`: Time-averaged flow quantities on the surface boundary (approx. 13 GB compressed, 21 GB unzipped).\n* `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`).\n* `volume_geo_LHCi_AoA_j.vtu.tgz`: Time-averaged flow quantities within the domain volume (approx. 23 GB compressed, 86 GB unzipped).\n* `geo_LHCi_AoA_j.stl`: Surface mesh (tris) of the geometry (approx. 197 MB).\n* `geo_LHCi_AoA_j.stp`: Surface geometry definition (approx. 48 MB).\n* `force_mom_geo_LHCi_AoA_j.csv`: Time-averaged drag, lift, moment, and pressure/viscous coefficients. Six documented cases use mean coefficients integrated from the released time-averaged surface fields.\n* `geo_values_geo_LHCi_AoA_j.csv`: Reference geometry values used to define the geometry via the DoE method.\n* `ref_values_geo_LHCi_AoA_j.csv`: Reference values such as area, Q, and AoA.\n* `img_wss_LHCi_AoA_j.png`: Visual rendering of the wall shear stress/skin-friction on the aircraft surface.\n* `plot_CD_geo_LHCi_AoA_j.png`: Convergence plot of the Drag Coefficient over time.\n* `plot_CL_geo_LHCi_AoA_j.png`: Convergence plot of the Lift Coefficient over time.\n* `plot_CM_geo_LHCi_AoA_j.png`: Convergence plot of the Pitching Moment Coefficient over time.\n\nIn addition to the files per run folder, there are also general summary files in the root directory:\n* `geo_values_all.csv`: Reference geometry values used to define the ge","encoding":"utf8","key":"README.md","name":"README.md","size":14573,"truncated":true},{"bytes":"# HiLiftAeroML native-volume invalid zero-fill rows\n\nThis directory identifies native-volume rows that should be ignored when\nusing the released volume field data. Every exported flow and statistics field\nwas verified to be exactly zero on each listed row. This includes stored\npressure, absolute temperature, and density, so the rows do not contain a\nphysically valid CFD solution state. The all-zero pattern is consistent with\nzero-filled placeholder or missing data; the files identify the affected rows\nbut do not assign a specific cause to the zero fill.\n\n## Files\n\n- `volume_zero_fill_raw_point_ids_all1800.jsonl` lists the case-local raw row\n  IDs for all 1,800 cases in human-readable ASCII.\n- `volume_zero_fill_raw_point_xyz_all1800.csv` provides the corresponding raw\n  native VTU coordinates as a human-readable diagnostic table with columns\n  `case_id,raw_point_id,x,y,z`.\n- `volume_zero_fill_raw_point_ids_all1800_manifest.json` documents how the rows\n  were identified, binds every source archive and both public data files by\n  SHA-256, and records integrity checks.\n- `SHA256SUMS` contains the publication checksums.\n\n## What the list records\n\nFor each case, the listed IDs are zero-based indices in the stored\n`PointData` row order of `volume_<case>.vtu`. They identify rows where the raw\nstored Float32 `avg(P)` is exactly numeric zero, including either sign of\nzero, using no tolerance. The locations were identified before Cp conversion,\nnormalization, or nondimensionalization and were not derived from model\npredictions. This raw-field condition is not equivalent to physical\n`Cp == 0`.\n\nThe initial `avg(P) == 0.0` test was used to find candidate rows. A subsequent\ncheck verified that all exported fields listed in the manifest are exactly\nzero on every candidate row. The published IDs are therefore the rows to\nignore; legitimate physical points are not selected merely because a derived\nquantity such as `Cp` is zero.\n\nAcross all 1,800 cases there are 419,416,158,837 raw volume rows and 1,232,817\nlisted zero-fill rows. The list is nonempty for 1,768 cases and empty for 32\ncases.\n\n## Reading one case\n\n```python\nimport json\n\nwith open(\n    \"data_quality/volume_zero_fill_raw_point_ids_all1800.jsonl\",\n    encoding=\"ascii\",\n) as stream:\n    for line in stream:\n        record = json.loads(line)\n        if record[\"case_id\"] == case_id:\n            zero_fill_raw_point_ids = record[\"zero_fill_raw_point_ids\"]\n            break\n    else:\n        raise KeyError(case_id)\n```\n\nThe file has exactly one compact JSON object per line, ordered by\n`case_index`; line number is therefore `case_index + 1`. IDs within a case are\nsorted and unique. A consumer processing a chunk can use these IDs to locate\nand omit the listed rows that fall inside that chunk.\n\n## Coordinate convenience table\n\nThe CSV contains one header and exactly 1,232,817 data rows, ordered by\n`case_index` and then `raw_point_id`. Its `x`, `y`, and `z` values are copied\nfrom `Points[raw_point_id]` in the native volume VTU without transformation or\nunit conversion. The source coordinates are Float32; each value is written\nwith nine significant decimal digits and verified to round-trip to the exact\nstored Float32 value.\n\nCoordinates are descriptive convenience data, not identifiers. Use the pair\n`(case_id, raw_point_id)` for exact joins. Do not use coordinate rounding,\nnearest-neighbour matching, or a geometric tolerance to redefine row\nidentity.\n\n## Source provenance\n\nThe manifest pins `nvidia/HiLiftAeroML` at immutable revision\n`1c266d3869bc2968ff97d2107c9c3919be03ed32`. For every case it records the\nrepository-relative volume-archive path, byte size, Hugging Face LFS SHA-256,\nand extracted VTU member name. The row-ID space is the named VTU obtained by\nlossless extraction of that pinned archive.\n\n## Scope\n\nThis metadata applies only to the listed native-volume `PointData` rows. It\ndoes not identify or describe any surface rows.\n\nSee the repository's root dataset card for licensing and citation details.\n","encoding":"utf8","key":"data_quality/README.md","name":"README.md","size":4014,"truncated":false},{"bytes":"# HiLiftAeroML Dataset Splits\n\nDeterministic train/val/test splits for the [HiLiftAeroML](https://huggingface.co/datasets/nvidia/HiLiftAeroML) dataset. Covers the **1800 complete LHC cases**: 180 geometries × 10 angles of attack (4-22 deg).\n\n## Splits at a glance\n\n| Split | Type | Train | Val | Test | What it tests |\n|----------------|-------|----:|---:|----:|------------------------------|\n| `full` | In-dist | 1260 | 180 | 360 | Baseline: random case-level split |\n| `medium` | In-dist | 510 | 180 | 360 | Intermediate data efficiency between `full` and `scarce` |\n| `scarce` | In-dist | 210 | 180 | 360 | Data efficiency (1/6 of `full` training data) |\n| `super_scarce` | In-dist | 35 | 180 | 360 | Extreme data efficiency (1/36 of `full` training data) |\n| `geometry` | In-dist | 1260 | 180 | 360 | Generalization to unseen geometries |\n| `geometry_medium` | In-dist | 510 | 180 | 360 | Unseen-geometry generalization from 51 training geometries |\n| `geometry_scarce` | In-dist | 210 | 180 | 360 | Unseen-geometry generalization from 21 training geometries |\n| `geometry_super_scarce` | In-dist | 40 | 180 | 360 | Unseen-geometry generalization from 4 training geometries |\n| `single_aoa_4` | In-dist | 126 | 18 | 36 | Geometry generalization at 4 deg (pre-stall) |\n| `single_aoa_12` | In-dist | 126 | 18 | 36 | Geometry generalization at 12 deg (mid-range) |\n| `single_aoa_22` | In-dist | 126 | 18 | 36 | Geometry generalization at 22 deg (post-stall) |\n| `aoa` | OOD | 788 | 112 | 900 | AoA extrapolation: low AoA → high AoA |\n| `deflection` | OOD | 1260 | 180 | 360 | Geometry extrapolation: low → high deflection |\n| `stall` | OOD | 942 | 135 | 723 | Flow regime: pre-stall → post-stall |\n\n**Difficulty and data-efficiency ladders:**\n\n- Case-level data efficiency: `full` < `medium` < `scarce` < `super_scarce` (1260 / 510 / 210 / 35 training cases, same val/test)\n- Geometry data efficiency: `geometry` < `geometry_medium` < `geometry_scarce` < `geometry_super_scarce` (126 / 51 / 21 / 4 training geometries, all 10 AoAs per geometry, same val/test)\n- Geometry generalization: `full` < `geometry` < `deflection`\n\n## Which split should I use?\n\n- **Simple baseline** → `full`\n- **Geometry generalization** (primary ML challenge) → `geometry`\n- **Hardest OOD challenge** → `deflection` or `stall`\n- **AoA extrapolation** → `aoa`\n- **Case-level data efficiency study** → compare `super_scarce`, `scarce`, `medium`, and `full` (same val/test)\n- **Geometry data efficiency study** → compare `geometry_super_scarce`, `geometry_scarce`, `geometry_medium`, and `geometry` (same held-out geometries)\n- **Single-AoA evaluation** → `single_aoa_4`, `single_aoa_12`, `single_aoa_22`\n\n## Design principles\n\n- **Val is always in-distribution with train.** For OOD splits, the val set is drawn from the training-side population, never from the OOD test region. Hyperparameter tuning never sees out-of-distribution data.\n- **Consistent ratios.** Splits with random/ranked geometry selection use 80/20 test fraction (70/10/20 overall). The `aoa` and `stall` splits have ratios dictated by their physics-based boundaries.\n- **Shared held-out geometries.** The `geometry*` and `single_aoa_*` splits use the same 18 val and 36 test geometries, enabling direct comparison across training-set sizes and AoA regimes.\n- **Test set integrity.** The test set should not be used to inform any training or hyperparameter decisions.\n\n## Split details\n\n### `full`\n\nRandom 70/10/20 partition of all 1800 cases. A given geometry may appear at different AoA values in train, val, and test. The model may have seen the same geometry at other angles of attack.\n\n### `medium`, `scarce`, and `super_scarce`\n\nSame val/test as `full`, but with nested random subsamples of `full_train`. `medium` contains 510 cases, near the geometric midpoint between `full` (1260) and `scarce` (210); `scarce` uses 1/6 of `full_train` (210 cases), and `super_scarce` uses 1/36 (35 cases). A single deterministic priority ordering defines every level, so `super_scarce_train` ⊂ `scarce_train` ⊂ `medium_train` ⊂ `fu","encoding":"utf8","key":"splits/README.md","name":"README.md","size":7261,"truncated":true},{"bytes":"{\n    \"full_train\": [\n        \"geo_LHC001_AoA_4\",\n        \"geo_LHC001_AoA_6\",\n        \"geo_LHC001_AoA_10\",\n        \"geo_LHC001_AoA_12\",\n        \"geo_LHC001_AoA_14\",\n        \"geo_LHC001_AoA_18\",\n        \"geo_LHC002_AoA_4\",\n        \"geo_LHC002_AoA_8\",\n        \"geo_LHC002_AoA_10\",\n        \"geo_LHC002_AoA_12\",\n        \"geo_LHC002_AoA_14\",\n        \"geo_LHC002_AoA_16\",\n        \"geo_LHC002_AoA_22\",\n        \"geo_LHC003_AoA_6\",\n        \"geo_LHC003_AoA_12\",\n        \"geo_LHC003_AoA_14\",\n        \"geo_LHC003_AoA_16\",\n        \"geo_LHC003_AoA_18\",\n        \"geo_LHC003_AoA_20\",\n        \"geo_LHC003_AoA_22\",\n        \"geo_LHC004_AoA_4\",\n        \"geo_LHC004_AoA_10\",\n        \"geo_LHC004_AoA_12\",\n        \"geo_LHC004_AoA_14\",\n        \"geo_LHC004_AoA_16\",\n        \"geo_LHC004_AoA_18\",\n        \"geo_LHC004_AoA_20\",\n        \"geo_LHC004_AoA_22\",\n        \"geo_LHC005_AoA_6\",\n        \"geo_LHC005_AoA_8\",\n        \"geo_LHC005_AoA_12\",\n        \"geo_LHC005_AoA_14\",\n        \"geo_LHC005_AoA_16\",\n        \"geo_LHC005_AoA_20\",\n        \"geo_LHC005_AoA_22\",\n        \"geo_LHC006_AoA_4\",\n        \"geo_LHC006_AoA_6\",\n        \"geo_LHC006_AoA_8\",\n        \"geo_LHC006_AoA_10\",\n        \"geo_LHC006_AoA_12\",\n        \"geo_LHC006_AoA_14\",\n        \"geo_LHC006_AoA_18\",\n        \"geo_LHC006_AoA_20\",\n        \"geo_LHC006_AoA_22\",\n        \"geo_LHC007_AoA_4\",\n        \"geo_LHC007_AoA_6\",\n        \"geo_LHC007_AoA_8\",\n        \"geo_LHC007_AoA_10\",\n        \"geo_LHC007_AoA_14\",\n        \"geo_LHC007_AoA_16\",\n        \"geo_LHC007_AoA_18\",\n        \"geo_LHC008_AoA_6\",\n        \"geo_LHC008_AoA_10\",\n        \"geo_LHC008_AoA_12\",\n        \"geo_LHC008_AoA_16\",\n        \"geo_LHC008_AoA_18\",\n        \"geo_LHC008_AoA_20\",\n        \"geo_LHC008_AoA_22\",\n        \"geo_LHC009_AoA_4\",\n        \"geo_LHC009_AoA_6\",\n        \"geo_LHC009_AoA_8\",\n        \"geo_LHC009_AoA_10\",\n        \"geo_LHC009_AoA_12\",\n        \"geo_LHC009_AoA_14\",\n        \"geo_LHC009_AoA_16\",\n        \"geo_LHC010_AoA_4\",\n        \"geo_LHC010_AoA_6\",\n        \"geo_LHC010_AoA_8\",\n        \"geo_LHC010_AoA_12\",\n        \"geo_LHC010_AoA_14\",\n        \"geo_LHC010_AoA_16\",\n        \"geo_LHC010_AoA_20\",\n        \"geo_LHC010_AoA_22\",\n        \"geo_LHC011_AoA_4\",\n        \"geo_LHC011_AoA_6\",\n        \"geo_LHC011_AoA_10\",\n        \"geo_LHC011_AoA_12\",\n        \"geo_LHC011_AoA_14\",\n        \"geo_LHC011_AoA_16\",\n        \"geo_LHC011_AoA_18\",\n        \"geo_LHC011_AoA_20\",\n        \"geo_LHC011_AoA_22\",\n        \"geo_LHC012_AoA_4\",\n        \"geo_LHC012_AoA_6\",\n        \"geo_LHC012_AoA_8\",\n        \"geo_LHC012_AoA_12\",\n        \"geo_LHC012_AoA_14\",\n        \"geo_LHC012_AoA_20\",\n        \"geo_LHC013_AoA_4\",\n        \"geo_LHC013_AoA_8\",\n        \"geo_LHC013_AoA_10\",\n        \"geo_LHC013_AoA_12\",\n        \"geo_LHC013_AoA_14\",\n        \"geo_LHC013_AoA_16\",\n        \"geo_LHC013_AoA_18\",\n        \"geo_LHC013_AoA_20\",\n        \"geo_LHC013_AoA_22\",\n        \"geo_LHC014_AoA_4\",\n        \"geo_LHC014_AoA_6\",\n        \"geo_LHC014_AoA_8\",\n        \"geo_LHC014_AoA_10\",\n        \"geo_LHC014_AoA_22\",\n        \"geo_LHC015_AoA_6\",\n        \"geo_LHC015_AoA_8\",\n        \"geo_LHC015_AoA_10\",\n        \"geo_LHC015_AoA_14\",\n        \"geo_LHC015_AoA_18\",\n        \"geo_LHC015_AoA_20\",\n        \"geo_LHC016_AoA_6\",\n        \"geo_LHC016_AoA_8\",\n        \"geo_LHC016_AoA_12\",\n        \"geo_LHC016_AoA_14\",\n        \"geo_LHC016_AoA_16\",\n        \"geo_LHC016_AoA_20\",\n        \"geo_LHC016_AoA_22\",\n        \"geo_LHC017_AoA_6\",\n        \"geo_LHC017_AoA_8\",\n        \"geo_LHC017_AoA_10\",\n        \"geo_LHC017_AoA_12\",\n        \"geo_LHC017_AoA_14\",\n        \"geo_LHC017_AoA_18\",\n        \"geo_LHC017_AoA_20\",\n        \"geo_LHC018_AoA_4\",\n        \"geo_LHC018_AoA_6\",\n        \"geo_LHC018_AoA_10\",\n        \"geo_LHC018_AoA_12\",\n        \"geo_LHC018_AoA_14\",\n        \"geo_LHC018_AoA_16\",\n        \"geo_LHC018_AoA_18\",\n        \"geo_LHC018_AoA_20\",\n        \"geo_LHC018_AoA_22\",\n        \"geo_LHC019_AoA_4\",\n        \"geo_LHC019_AoA_6\",\n        \"geo_LHC019_AoA_8\",\n        \"geo_LHC019_AoA_10\",\n        \"geo_LHC019_AoA_12\",\n        \"geo_LHC019_AoA_14\",\n        \"geo_LHC019_AoA_16\",\n        \"geo_LHC019_AoA_18\",\n        \"geo_LHC019_AoA_20\",\n        \"geo_LHC019_AoA_22\",\n        \"geo_LHC020_AoA_4\",","encoding":"utf8","key":"splits/manifest.json","name":"manifest.json","size":411540,"truncated":true}],"sourceObjectKey":"data_quality/volume_zero_fill_raw_point_ids_all1800.jsonl","starterCell":"from datasets import load_dataset\nds = load_dataset(\"nvidia/HiLiftAeroML\")\nprint(ds)","starterSamplePath":"data_quality/volume_zero_fill_raw_point_ids_all1800.jsonl","totalObjects":23423},"summary":"High-fidelity computational fluid dynamics dataset of a high-lift transport aircraft, built for training and testing AI surrogate models in aerospace aerodynamics.","title":"HiLiftAeroML: Wall-Modeled LES of the NASA CRM High-Lift Aircraft for AI Surrogates","updatedAt":"2026-11-18T03:59:41.143Z","visibility":"public","websiteUrl":"https://huggingface.co/datasets/nvidia/HiLiftAeroML"}}