High-Reynolds Turbulent Datasets

• High-Reynolds turbulent datasets are comprehensive resources that capture the complex, multiscale physics and universal scaling laws of turbulent flows at Re numbers typically exceeding 10⁴.
• They support validation of numerical methods like DNS, LES, and RANS and enable the development of turbulence models and machine learning surrogates with detailed statistical and full-field data.
• Applications include simulation benchmarking, flow control, and fundamental analysis across canonical, free-shear, and complex geometries, utilizing experimental and hybrid computational techniques.

High-Reynolds-number turbulent datasets are comprehensive data resources that capture the complex, multiscale physics of turbulent flows at Reynolds numbers far exceeding transitional or low-Reynolds regimes. These datasets are foundational for the characterization of universal scaling properties, the validation of numerical methods (DNS, LES, RANS), the development of turbulence models, as well as the training and benchmarking of machine learning-based surrogates and closures. Contemporary high-Reynolds datasets span canonical wall-bounded and free-shear flows, experimental measurements, large-eddy and Reynolds-averaged simulations, and application-specific scenarios such as vehicle wakes and roughness transitions. Encompassing both statistical moments and full field information, these datasets enable rigorous interrogation of turbulence structure, non-equilibrium dynamics, and scaling laws throughout the high-Re regime.

1. Definitions, Importance, and Scope

The defining parameter for high-Reynolds-number turbulence is the Reynolds number: $\operatorname{Re} = UL/\nu$, where $U$ is a characteristic velocity, $L$ is a lengthscale (e.g., pipe diameter, boundary layer thickness), and $\nu$ is the kinematic viscosity. Flows with $\operatorname{Re} \gtrsim 10^4$ in wall units ($Re_\tau$), or up to order $10^8$ as in the largest published computational datasets, are considered high-Re and exhibit strongly separated scales, inertial subranges, and pronounced large-scale motions.

High-Re turbulent datasets are required to:

• Capture universal inertial range features, such as the $k^{-5/3}$ and $k^{-1}$ velocity spectra observed over multiple decades of scale (Nagib et al., 9 Nov 2025, Yao et al., 2022)
• Resolve the emergence of logarithmic regions in mean and fluctuating statistics, including velocity, normal stresses, and wall-shear fluctuations
• Support direct validation of turbulence models and wall laws relevant to engineering applications at real-world Reynolds numbers
• Enable machine learning models to generalize beyond low-Re regimes, essential for robust surrogate modeling and predictive flow control (Cooper-Baldock et al., 1 Feb 2026)

Comprehensive high-Re datasets exist for canonical flows (pipe, channel, boundary layer), free-shear flows (jets, wakes), practical geometries (vehicle recovery wakes), and flows with wall condition changes (rough-to-smooth transitions).

2. Experimental and Computational Techniques

High-Re datasets are realized via large-scale laboratory experiments, high-fidelity direct numerical simulation (DNS), large-eddy simulation (LES), hybrid methods, and RANS-based parametrizations, each with specific strengths and limitations.

Experimental Datasets:

• Facilities such as CICLoPE (closed-loop pipe, $Re_\tau$ up to 50,000, precision Pitot and hot-wire, $U$0 precision ±0.2%) provide long-asymptotic fully-developed pipe flow data with sub-micron wall roughness (Nagib et al., 9 Nov 2025).
• Boundary-layer wind tunnels (Melbourne HRNBLWT) deliver controlled zero- and adverse-pressure-gradient wall-bounded flows, and enable systematic study of rough-to-smooth transitions and TNTI dynamics (Li et al., 2023, Marusic et al., 2024).
• Advanced PIV systems deliver three-component, high dynamic-range fields in the outer regions required for TNTI and conditional statistics at $U$1 (Marusic et al., 2024).

Direct Numerical Simulation (DNS):

• OpenPIPEFLOW and NEK5000 have generated smooth-pipe datasets up to $U$2, with grid spacings as fine as $U$3 and substantial averaging statistics (Yao et al., 2022).
• Compressible pipe DNS at moderate Re up to $U$4 enables study of compressibility effects and similarity transformations (Modesti et al., 2018).

LES and Hybrid Approaches:

• LES of subsonic jets at $U$5 and airfoil wakes (Towne et al., 2022)
• Surface-sampled LES–QDNS couples coarse-grid LES with near-wall quasi-DNS to efficiently reach $U$6 with detailed statistics up to the buffer and log-layer (Sandham et al., 2017).

Parametric RANS & Application-Specific CFD:

• WAKESET provides $U$7 high-fidelity RANS cases (augmented to 4,364 instances), up to $U$8, resolving complex marine hydrodynamics (Cooper-Baldock et al., 1 Feb 2026).

3. Statistical Content, Scaling Laws, and Universal Features

High-Re turbulent datasets consistently reveal the emergence and stabilization of universal scaling features:

Mean Flow and Law of the Wall:

• The mean centerline velocity in pipe flow approaches $U$9 with $L$0 beyond $L$1 (Nagib et al., 9 Nov 2025).
• For smooth pipes, the best-fit log-law constant is $L$2 in DNS, with log-indicator function $L$3 up to $L$4 (Yao et al., 2022).

Normal Stress and Fluctuations:

• Streamwise turbulence intensity on the pipe centerline exhibits a plateau of $L$5 and $L$6 at $L$7 (Nagib et al., 9 Nov 2025).
• Boundary layer and pipe data confirm the modified Townsend–Perry attached-eddy model:

$L$8

with $L$9 and $\nu$0 across facilities. An outer-peak/plateau appears at $\nu$1 for $\nu$2 (Laval et al., 2018).

Spectral Content:

• One-dimensional pre-multiplied spectra demonstrate $\nu$3 inertial subrange scaling over seven decades in CICLoPE (Nagib et al., 9 Nov 2025), and $\nu$4 region in both DNS and experiments for $\nu$5 (Yao et al., 2022).
• Spectral collapse in compressible pipe DNS using effective eddy length scaling $\nu$6 (Modesti et al., 2018).

Higher-Order Statistics:

• Skewness $\nu$7 and kurtosis $\nu$8 of streamwise centerline velocity in pipes show no Reynolds number dependence from $\nu$9 to $\operatorname{Re} \gtrsim 10^4$0 (Nagib et al., 9 Nov 2025).

4. Dataset Structure, Variables, and Access

Representative datasets exhibit standardized structure and content:

Dataset	Reynolds Number(s)	Format/Content
WAKESET	$\operatorname{Re} \gtrsim 10^4$1	1,091 RANS cases, 4,364 instances, 128$\operatorname{Re} \gtrsim 10^4$2 volumes, 512$\operatorname{Re} \gtrsim 10^4$3 slices, Python/NumPy loaders, open-source on HuggingFace (Cooper-Baldock et al., 1 Feb 2026)
CICLoPE	$\operatorname{Re} \gtrsim 10^4$4	UCL, $\operatorname{Re} \gtrsim 10^4$5, spectra, higher moments (Nagib et al., 9 Nov 2025)
PIPE DNS	$\operatorname{Re} \gtrsim 10^4$6	HDF5, full 3D fields, mean, budgets, spectra. Public repository (Yao et al., 2022)
Surface-LES-QDNS	$\operatorname{Re} \gtrsim 10^4$7	LES fields, embedded near-wall QDNS, mean/fluctuation statistics, spectra, code/scripts (Sandham et al., 2017)
TBL Datasets	$\operatorname{Re} \gtrsim 10^4$8	Hot-wire, PIV (planar/volumetric), velocity profiles, variance, TNTI conditionality (Laval et al., 2018, Marusic et al., 2024)

WAKESET uniquely provides a 480 GB open-access set of highly augmented, variable-parameter turbulent wake fields suitable for machine learning benchmarks (Cooper-Baldock et al., 1 Feb 2026). CICLoPE and Superpipe archives contain complete profiles and spectra available on request or via direct institutional contact (Nagib et al., 9 Nov 2025). High-fidelity DNS data are distributed via public DOIs and enable rigorous comparative studies (Yao et al., 2022).

5. Applications and Benchmarking

Application domains include:

• Simulation Benchmarking: Wall laws, turbulence closures, subgrid models, wall-modelled LES, and hybrid approaches require accurate high-Re data for calibration and validation. The surface-sampled LES–QDNS dataset is specifically designed to validate wall-models under non-equilibrium conditions (Sandham et al., 2017).
• Physics-Informed Machine Learning: Training of generative models, surrogates, conditional flow field predictors, and turbulence closure representations is directly enabled by datasets like WAKESET—metrics used include PSNR, SSIM, FID, and physics-informed energy error $\operatorname{Re} \gtrsim 10^4$9 (Cooper-Baldock et al., 1 Feb 2026).
• Flow Control and Reduced-Order Modeling: Public jet, boundary layer, and airfoil-wake datasets are foundational for comparing linear reduced-complexity models (SPOD, DMD, resolvent), sparse-sensing, and modal-control strategies (Towne et al., 2022).
• Turbulence Structure and Scaling Law Analysis: Universal features such as the emergence of log-law plateaus, $Re_\tau$0 spectra, and outer-peak dynamics are systematically interrogated with comprehensive high-Re datasets (Yao et al., 2022, Laval et al., 2018).
• Non-Equilibrium and Transitional Effects: Rough-to-smooth experiments elucidate the non-equilibrium recovery of wall stress and mean-velocity blending, challenging two-log-law internal boundary layer models (Li et al., 2023).

6. Comparison Among Data Sources and Limitations

Across methodologies, high-Re datasets reveal converging but not identical results:

• Spectral DNS consistently predict slightly higher friction factors, wall stress fluctuations, and velocity intensities than lower-order finite-difference DNS; experimental plateau values most closely match high-order DNS (Yao et al., 2022, Nagib et al., 9 Nov 2025).
• WAKESET exceeds all prior published machine learning–focused CFD datasets in both Reynolds number ($Re_\tau$1 vs $Re_\tau$2) and the breadth of operational parameterization (speeds and yaw angles) (Cooper-Baldock et al., 1 Feb 2026).
• Laboratory datasets at $Re_\tau$3 are essential to observe true high-Re asymptotic behavior; below this threshold, scaling quantities may not be fully developed (Nagib et al., 9 Nov 2025).
• Maximum Re in DNS is currently $Re_\tau$4 for incompressible, $Re_\tau$5 for compressible pipe; ultra-high Re ($Re_\tau$6) is currently accessible only to experiments or RANS/LES parameterizations (Yao et al., 2022, Modesti et al., 2018).
• Experimental datasets are susceptible to wire length/attenuation corrections (NSTAP vs hot-wire), near-wall reflection contamination (PIV), and precise velocity/friction velocity definitions, all requiring careful data handling for cross-comparison (Laval et al., 2018).

7. Emerging Directions and Availability

Advances in computational resources, hybrid numerical methods, and integrated machine learning promise further extension of high-Re datasets:

• Public, permissively licensed data (as in WAKESET and the UM Database) is increasingly available, with code and environments (Docker files) for replicability (Cooper-Baldock et al., 1 Feb 2026, Towne et al., 2022).
• Hybrid approaches (LES-QDNS) enable parameteric studies into $Re_\tau$7 at moderate cost, serving as a bridge between DNS and RANS/LES (Sandham et al., 2017).
• Datasets optimized for machine learning now provide not only fields, but also standardized train/validation/test splits and open performance baselines (GANs, cDCGAN, WGAN-GP) (Cooper-Baldock et al., 1 Feb 2026).
• Multi-physics expansions (compressible, reactive, scalar transport) and non-equilibrium surface conditions (roughness, pressure gradients, TNTI dynamics) are now represented in emerging datasets (Marusic et al., 2024, Li et al., 2023).

A plausible implication is that the continued growth in dataset scale, diversity, and open availability will accelerate the development of data-driven turbulence models and provide robust testbeds for physics-based and hybrid approaches spanning the full range of turbulent flow regimes.