This dataset is released under CC BY-NC-SA 4.0. Research and evaluation use is explicitly welcome — academic research, internal engineering evaluation, benchmarking, and educational use are all permitted. Productization is not permitted: you may not incorporate this dataset, in whole or in part, into a commercial product or service, or use it to train or validate models that are sold, licensed, or offered as part of a commercial offering. Contact the maintainers if you want to discuss a separate commercial license.

nTop × CoreWeave Drone Planform LBM-LES DOE Sweep

Dataset Summary

This dataset is the tabular output of a large-scale Latin Hypercube Sampling (LHS) Design of Experiments (DOE) sweep over fixed-wing drone planform geometries, evaluated with nTop Fluids — a GPU-native Lattice Boltzmann Method (LBM) solver using Wall-Modeled Large Eddy Simulation (WMLES) turbulence treatment — and executed on CoreWeave's GPU infrastructure. It was produced as a proof-of-concept (POC) collaboration between nTop and CoreWeave, jointly demonstrating GPU-native CFD throughput at scale and generating a design-space dataset intended as a foundation for surrogate-based multidisciplinary design optimization (MDO).

The target platform is a Group 3 long-endurance flying-wing UAS with a NACA 0008 airfoil, a fixed span of 3,657.6 mm, flown at cruise conditions of Mach 0.21 at 30,000 ft (ISA standard atmosphere: T = 228.71 K, ρ = 0.4589 kg/m³). The six free geometric parameters describe the leading- and trailing-edge sweep of a two-panel planform at fixed span; angle of attack (AoA) was swept independently at five discrete values.

This release hosts the full per-run output set produced by the simulation campaign, not only the aggregated scalar table. For each of the approximately 11,948–11,950 completed (geometry, AoA) evaluations, the dataset provides: aerodynamic scalar outputs (C_L, C_D, L/D, pitching moment, lift, drag, reference area, and two structural-weight proxies); spatially resolved surface pressure and force distributions; and a volumetric flow-field result in VTI format containing pressure and velocity fields. Aggregate dataset size is approximately 13–14 TB. A consolidated scalar-only CSV (results.csv), aggregating the integral results across all runs, is provided at the dataset root for users who only need the tabular aerodynamic outputs and do not require the per-run volumetric or surface fields.

Scope & Character of This Dataset — Read This First

This is not a cleaned-up academic benchmark, and it isn't meant to become one. It is the direct, one-time output of a time-boxed industrial DOE campaign — roughly 40 hours of wall-clock execution against a fixed compute allocation — run to demonstrate throughput and infrastructure at scale, not to produce a polished, ready-to-train dataset. There is no planned "v2 cleanup" pass beyond the outlier-tier scheme already documented below. What you are looking at is what came off the cluster.

Concretely, that means:

results.csv is raw, not pre-filtered. It contains all submitted evaluations, including the T1 hard-artefact rows, the T2/T3 statistically flagged rows, and the failed jobs recorded via status. The three-tier scheme in Data Quality & Outlier Filtering below is a recommended filter for you to apply, not a cleaning pass that has already been done to the file you're downloading.
Some runs did not succeed. A small fraction of jobs (see status, and the TBB-related restarts noted under Known Limitations) did not produce complete output. Don't assume every {batch_id}-job-{N}/ directory contains a full, valid set of files — check status and the presence of expected files before processing in bulk.
Solver fidelity is coarse by design, calibrated for design-space coverage across thousands of geometries rather than per-case numerical accuracy. This is not a verification-grade CFD dataset.
The directory structure is not flattened. Per-run files sit exactly where the original CoreWeave execution wrote them — nested under {batch_id}/{batch_id}-job-{N}/inputs/ and outputs/ — mirroring the GCS layout used during the campaign itself, rather than being reorganized into a single tidy per-modality folder tree. Working with the full per-run assets means walking that structure, not listing a flat directory.

We're calling this out deliberately rather than apologetically. The value of this dataset is partly because it looks like this: it's representative of what a real, accelerated-compute CFD DOE campaign actually produces under time pressure — artefacts, partial failures, coarse resolution, and all — rather than a dataset that has already had those characteristics engineered out of it. If you're building tooling for outlier detection, robust surrogate training, or data pipelines against large-scale simulation campaigns, treat that rawness as the point, not as a defect to route around. If you just want clean scalar numbers to plot, apply the T1 filter first and go from there.

Motivation

Systematic CFD-based characterization of a parametric design space is operationally expensive even when the underlying geometry parameterization is low-dimensional, because it requires a pipeline capable of generating geometry variants, dispatching simulation jobs, and collecting outputs at scale without per-case human intervention. This dataset was generated to (1) demonstrate nTop Fluids' throughput at a scale that makes the case concretely rather than theoretically, (2) validate CoreWeave's Kubernetes-based GPU infrastructure as a substrate for sustained, high-concurrency CFD DOE workloads, and (3) produce a design-space dataset of sufficient scale and geometric diversity to support downstream Gaussian Process / Bayesian Optimization surrogate work. The effort is framed, with appropriate caveats, against NASA's CFD Vision 2030 Grand Challenge Problem 3 (large-scale ensemble simulation); it demonstrates meaningful throughput toward that benchmark without claiming to satisfy its full requirements (single-engineer interpretation and routine repeatability were not independently established here).

Dataset Structure

Instances

Each instance is one CFD evaluation: a unique planform geometry at one of five angles of attack.

Unique geometries sampled: 2,400 (post-feasibility-filtering)
AoA levels per geometry: 2°, 4°, 6°, 8°, 10°
Total submitted evaluations: 12,000 (2,400 × 5)
Completed simulations: approximately 11,948–11,950 (51 jobs required manual restart due to a known Threading Building Blocks (TBB) issue, occurring at a rate of roughly 0.4%; all recovered on restart)

Data Fields

Geometry inputs (sampled via scipy.stats.qmc.LatinHypercube, seed 2024, at fixed span of 3,657.6 mm):

Field	Description	Bounds
`loa_mm`	Root chord (length overall)	1,400.8 – 4,698.6 mm
`le_sweep_p1_deg`	Leading-edge sweep, inboard panel	0.03° – 54.97°
`le_sweep_p2_deg`	Leading-edge sweep, outboard panel	−7.98° – 50.00°
`te_sweep_p1_deg`	Trailing-edge sweep, inboard panel	−44.89° – 54.96°
`te_sweep_p2_deg`	Trailing-edge sweep, outboard panel	−49.98° – 49.95°
`panel_break_pct`	Spanwise panel break as fraction of half-span	0.250 – 0.700

Condition input:

Field	Description
`aoa_deg`	Angle of attack (2, 4, 6, 8, or 10°)

Aerodynamic outputs (from LBM surface pressure integration; scalar quantities only — this is not a vortex-lattice or panel-method result):

Field	Description
`C_L (1)`	Lift coefficient
`C_D (1)`	Drag coefficient
`L/D (1)`	Glide ratio (lift-to-drag)
`PM (J)`	Pitching moment
`Lift (N)`	Lift force
`Drag (N)`	Drag force
`ref_A (mm²)`	Reference planform area
`weight_composite (kg)`	Structural weight proxy, composite construction
`weight_metal (kg)`	Structural weight proxy, metal construction
`status`	Job completion status

Identifier & diagnostic fields:

Field	Description
`job_id`	Numeric identifier for the evaluation, joins to the Top-N / outlier-tier tables
`batch`	Batch identifier (e.g. `batch-20260428-162600`), corresponds to the `{batch_id}` path segment in the repository layout below
`batch_job_id`	Full job identifier within the batch, corresponds to the `{batch_id}-job-{N}` path segment; use this to locate a run's `inputs/` and `outputs/` files
`geom_id`	Convenience integer index over the 2,400 unique geometries (derived from `loa_mm`); not part of the original job submission
`runtime_s`	Wall-clock simulation runtime for the job, in seconds
`cell size (mm)`	LBM voxel cell size actually used for the run; retained per-row to support auditing whether statistically flagged (T2) rows correlate with mesh resolution rather than genuine geometric effects

Simulation was performed at a cell size (LBM voxel resolution) selected for rapid design-space exploration across 12,000 variants rather than maximum per-geometry accuracy; a separate forthcoming publication documents the grid-convergence study characterizing the accuracy–cost relationship of this solver across cell sizes.

Data Splits

This dataset does not define train/validation/test splits. It is a single design-of-experiments corpus intended as raw input to downstream surrogate model construction (e.g., Gaussian Process regression via BoTorch/OpenMDAO), and split strategy is left to the consumer.

Repository Layout

Per-run outputs are organized by batch and job, mirroring the structure used during the original CoreWeave campaign. The dataset root holds this README plus two dataset-level files, with per-run data nested below:

README.md
results.csv                     # consolidated scalar outputs + geometry inputs, all runs
ntop-file.ntop                  # shared nTop project file used to generate the geometries
{batch_id}/
  {batch_id}-job-{N}/
    inputs/
      input_template.json       # the six geometry parameters + AoA for this run
    outputs/
      model.implicit            # implicit solid model, for reprocessing with nTop Core
      surface_mesh.obj          # surface mesh — visualization only, NOT the simulation mesh
      pressure-on-surface.csv   # spatially resolved surface pressure distribution
      forces-on-surface.csv     # spatially resolved surface force distribution
      cfd-result.vti            # volumetric 3D pressure + velocity flow field
      summary-dict.json         # per-run scalar outputs and run metadata

inputs/input_template.json records the exact geometry parameters and AoA used to generate that run, and is the authoritative per-run record of the six geometry inputs plus aoa_deg (rather than relying solely on the consolidated CSV for provenance).

outputs/model.implicit is an implicit solid representation of the geometry intended for reprocessing with nTop Core (e.g., regenerating a mesh at a different resolution, or extracting alternative surface representations); it is not itself a mesh.

outputs/surface_mesh.obj is provided for visualization purposes only. Because nTop Fluids is a Lattice Boltzmann solver operating directly on a volumetric grid, no surface mesh is used in, or produced by, the simulation itself — this .obj is a separate visualization-oriented export of the geometry and should not be interpreted as the discretization on which the CFD result was computed.

A consolidated results.csv, aggregating the summary-dict.json scalar fields (and the corresponding input_template.json parameters) across all runs (the fields listed under Data Fields above), is provided at the dataset root alongside this README, for convenience. The dataset root also includes ntop-file.ntop, the source nTop project file used to parametrically generate the planform geometries for this campaign — provided once at the dataset level rather than duplicated per run, since it is the shared generator rather than a per-geometry output. Given the aggregate size of the volumetric and per-surface data (~13–14 TB), users working primarily with scalar aerodynamic outputs should start from results.csv and only pull individual {batch_id}-job-{N}/ directories for runs whose flow fields, surface distributions, or per-run inputs they specifically need.

Data Quality & Outlier Filtering

The LBM solver's surface-pressure integration produces known artefacts at low angles of attack on the voxel meshes used here (approximately 10–20 cells across the chord), most visibly as small negative or anomalously large drag/efficiency values near zero lift. A three-tier filtering scheme is documented alongside the raw data and is recommended before using this dataset for statistical or surrogate-modeling work:

Tier	Criterion	Recommended action
T1	`C_D < 0` OR `C_L < 0` OR `\|L/D\| > 100`	Exclude — hard physical limits; unambiguous solver artefacts
T2	IQR (1.5×) and modified Z-score (`\|0.6745 × (x − median) / MAD\| > 3.5`) both flag the same output channel	Flag for review; excluded from conservative surrogate-training subsets
T3	Only one of the two statistical methods flags the row	Marginal; retained by default

Applying T1 alone removes 1,037 rows (≈8.7% of completed runs), concentrated almost entirely at 2° and 4° AoA. After T1 filtering, the maximum L/D in the dataset is a physically credible 52.8, versus an unfiltered maximum of approximately 1,429, which was entirely artefact-inflated. A second-order polynomial surrogate for L/D improves from R² = 0.21 (unfiltered) to R² = 0.63 (T1-filtered), and further to R² = 0.938 with additional refinement, underscoring that outlier handling materially affects any downstream modeling built on this data. Full filter implementation code is provided in the accompanying analysis documentation.

Note that two successive post-processing passes exist for this campaign, reflecting a refinement to the underlying turbulence-model configuration between passes; the revised pass reduces T1 artefact incidence substantially (from ≈8.7% to ≈0.2% of rows) but increases median C_D by roughly 52%, which roughly halves L/D values relative to the earlier pass. Users combining or comparing subsets of this data across passes should confirm which turbulence-model configuration produced the rows in question.

Data Collection & Infrastructure

Geometries were sampled via Latin Hypercube Sampling over the six-dimensional parameter space above, then passed through a post-sampling geometric feasibility filter (chord, taper, and sweep-crossing guards) that rejected approximately 20% of raw LHS candidates before the target of 2,400 valid geometries was reached. Each accepted geometry was crossed with the five-point AoA sweep to produce the full 12,000-run job matrix, submitted in chunks for cluster dispatch.

Simulation was performed using nTop Fluids (GPU-native LBM with WMLES turbulence treatment) on CoreWeave's Kubernetes infrastructure, provisioned across 35 nodes of NVIDIA RTX Pro 6000 Blackwell Server Edition GPUs (280 GPUs total, one simulation dispatched per GPU). Job orchestration, queuing, and progress tracking were handled by nTop's batch orchestrator; this campaign served as its largest real-world stress test to date. Per-job runtime is recorded (runtime_s) and is available for anyone auditing throughput or reproducing timing figures; the dashboard used for post-processing tracks median and P95 runtime alongside the failure and outlier-tier metrics.

Known Limitations

See also Scope & Character of This Dataset above — the points below are the specific technical consequences of this being a raw, one-time industrial sweep rather than a curated benchmark.

Solver fidelity is exploration-oriented, not verification-grade. The voxel cell size was selected to prioritize design-space coverage across thousands of geometries over per-case numerical accuracy. Absolute aerodynamic values should be treated as directionally informative for sensitivity and surrogate-training purposes rather than as validated point predictions; a separate grid-convergence study is forthcoming.
Low-AoA artefacts require filtering. See the T1/T2/T3 scheme above; using this dataset without applying at least T1 filtering will materially bias any downstream statistics or surrogate model.
Structural outputs are proxies, not simulated. weight_composite and weight_metal are parametric estimates, not outputs of coupled structural (FEM) simulation; structural analysis was explicitly descoped from this campaign due to the per-geometry meshing latency it would have introduced.
Aggregate size is large (~13–14 TB). Downloading the full repository, including per-run volumetric (cfd-result.vti) and surface (pressure-on-surface.csv, forces-on-surface.csv) data for all ~11,950 runs, requires substantial storage and bandwidth; most users will want to selectively pull per-run directories rather than clone the full dataset, and should start from the dataset-root results.csv for scalar-only workflows.
A small fraction of jobs required manual restart (≈0.4%, attributed to a known TBB threading issue), after which they completed without further incident; this does not materially affect the coverage or validity of the completed dataset.

Intended Uses

This dataset is intended to support:

Global sensitivity analysis and partial-dependence characterization of drone planform aerodynamics (Spearman rank correlation and Random Forest importance analyses have been performed and are documented separately)
Training of surrogate models (polynomial, Gaussian Process, or neural) for use in multidisciplinary design optimization loops
Pareto-front analysis trading aerodynamic efficiency against structural weight proxies
Benchmarking of GPU-native LBM-LES throughput for large-scale CFD DOE campaigns

It is not intended as a certified aerodynamic reference dataset for a specific vehicle program, given the fidelity caveats above. Productization is out of scope by license: this dataset may not be incorporated into a commercial product or service, or used to train or validate a model that is itself sold or commercially offered — see Licensing & Access below.

Acknowledgements

This dataset was produced through a collaboration between nTop and CoreWeave. The maintainers thank the CoreWeave infrastructure and project team for GPU compute access and cluster support, and the nTop engineering team for geometry automation, solver execution, and post-processing.

Citation

If you use this dataset, please cite the accompanying nTop × CoreWeave DOE retrospective and analysis report. A formal citation (paper, DOI, or preprint) will be added here upon publication.

Licensing & Access

This dataset is released under CC BY-NC-SA 4.0 (Attribution–NonCommercial–ShareAlike). In plain terms: research and evaluation use is welcome — academic research, internal engineering evaluation, benchmarking, coursework, and similar non-commercial use are all fine, with attribution, and any derivative dataset must be shared under the same terms. Productization is what's off the table: you may not fold this dataset into a commercial product or service, or use it to train or validate a model that is itself sold, licensed, or offered commercially. If you want to do that, contact the maintainers to discuss a separate commercial license.

Access is gated on Hugging Face — requesting access requires submitting basic contact details (name, company/institution, country, email, and intended use) and confirming the use is research/evaluation rather than productization. This is a platform access-control mechanism for tracking who is using the dataset and is separate from the license terms above.

Downloads last month: 32

Total file size:

11.6 TB