Sup3rWind: Multi-Scale Wind Modeling

• Sup3rWind is a research label denoting multiple wind modeling programs applied across atmospheric super-resolution, renewable energy downscaling, and turbulence synthesis.
• Its atmospheric implementations include a physics-informed SR-GAN that spatially upscales 3D wind fields and a deep-learning reanalysis-to-resource framework validated against observational data.
• In astrophysics, Sup3rWind models galactic superwinds and compact-object outflows, offering multiscale insights while highlighting context-specific challenges.

Sup3rWind is a polysemous research label applied to several wind-focused modeling programs. In current arXiv literature it denotes, most prominently, a physics-informed super-resolution generative adversarial network for three-dimensional atmospheric wind fields, a spatiotemporal downscaling framework and public wind-resource dataset derived from ERA5 and the WIND Toolkit, a superstatistical generator of intermittent wind fields constrained by sparse measurements, and multiple astrophysical formulations of galactic, cluster, AGN, and compact-object superwinds (Kurihana et al., 2023, Benton et al., 2024, Friedrich et al., 2022, Tanner et al., 2015, Danehkar et al., 2024).

1. Range of meanings

The available usages place Sup3rWind at the intersection of atmospheric modeling, renewable-energy resource assessment, stochastic turbulence synthesis, and astrophysical outflow theory. This suggests that the term functions less as a single canonical formalism than as a recurring label for multiscale wind reconstruction or transport problems in different domains.

Domain	Core meaning	Representative paper
Atmospheric super-resolution	Physics-informed SR-GAN for 3D winds	(Kurihana et al., 2023)
Renewable-energy downscaling	ERA5-to-2 km, 5-minute wind-resource product	(Benton et al., 2024)
Turbulence synthesis	Superstatistical wind fields from mast data	(Friedrich et al., 2022)
Galactic and cluster superwinds	Feedback-driven, cooling-regulated outflows	(Tanner et al., 2015, Danehkar et al., 2024, Wunsch et al., 2011)
Compact-object winds	SGXB wind accretion and pulsar-wind precursors	(Mellah et al., 2016, Mochol et al., 2013)

Two atmospheric usages are especially prominent. One reconstructs high-resolution three-dimensional winds from low-resolution numerical weather model outputs with explicit treatment of vertical convection (Kurihana et al., 2023). The other produces historically accurate, high-resolution wind-resource data from reanalysis, yielding a 24-year record for Ukraine, Moldova, and eastern Romania at 2-km spatial and 5-minute temporal resolution (Benton et al., 2024). A separate atmospheric usage treats Sup3rWind as a superstatistical extension of Gaussian inflow models, designed to reproduce intermittency and higher-order statistics from sparse measurements (Friedrich et al., 2022).

Astrophysical usages are broader. In that literature the label is attached to starburst-driven superbubbles, radiatively cooling galactic superwinds, strong-cooling super star cluster winds, AGN-driven super-winds, stationary halo-regulated superwinds, and compact-object wind problems (Tanner et al., 2015, Danehkar et al., 2023, Wunsch et al., 2011, Costa et al., 2020, Añorve-Zeferino et al., 2010, Mellah et al., 2016, Mochol et al., 2013).

2. Physics-informed super-resolution of three-dimensional wind fields

In the atmospheric super-resolution formulation, Sup3rWind is a preliminary super-resolution model for reconstructing high-resolution three-dimensional wind fields from low-resolution numerical weather outputs, with greenhouse-gas source tracing and monitoring as the stated application (Kurihana et al., 2023). The low-resolution input comprises the three wind components $u$, $v$, and $w$ on a three-dimensional grid with non-uniform vertical spacing, and the target is a horizontally super-resolved field. The generator learns

$$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$

with $C=3$, $H'=9H$, $W'=9W$, and $Z'=Z$. The stated upscaling factor is $\times 9$, corresponding to a horizontal resolution increase from 900 m to 100 m while preserving the number and spacing of vertical levels.

Its defining architectural component is the Pixel-Wise Self-Attention module. For each horizontal pixel $(i,j)$, the model restricts attention to the local vertical column $v$0. Self-attention is therefore used to encode cross-level couplings such as vertical convection, updrafts, and shear, after which 2D convolutional upsampling synthesizes fine horizontal structure. The vertical and horizontal operators are deliberately decoupled: attention handles the non-uniform vertical grid, while 2D convolutions target the horizontal-only enhancement and avoid the cost of full 3D convolution.

The model is described as physics-informed because pretraining includes a loss term that regularizes the self-attention map to capture the vertical convection process from the input wind data. The reported training objective combines adversarial, reconstruction, and attention-regularization terms, although the exact loss coefficients and the precise attention regularizer are not specified. No explicit enforcement of mass continuity or energy constraints is reported.

Evaluation is centered on spectral fidelity. Using fast Fourier transform analysis over 2,500 test samples, the study reports that SR-GAN outperforms the attention-only SR-PWA baseline for learning high-frequency domains, especially in vertical wind $v$1; that there is a performance trade-off in low- to medium-wave numbers when the high-frequency domain is learned well; that bicubic interpolation performs poorly in the high-frequency domain; and that a conventional residual CNN yields spectra similar to SR-PWA. Qualitative snapshots at the first vertical layer show that SR-PWA restores major wind structures and velocity intensities, while SR-GAN produces finer-scale features consistent with the high-resolution target. The computational cost of obtaining high-resolution wind fields is reduced by $v$2 relative to running high-resolution simulations, with the cost metric described as CPU time times number of cores (Kurihana et al., 2023).

3. Reanalysis-to-resource downscaling and the Ukraine dataset

A second atmospheric meaning of Sup3rWind is a generative, deep-learning-based spatiotemporal downscaling framework and dataset for renewable-energy resource assessment (Benton et al., 2024). Here the objective is historically accurate high-resolution wind resource data over long periods. The system is trained to map European Centre for Medium-Range Weather Forecasting Reanalysis version 5 inputs to high-resolution targets from the Wind Integration National Dataset, and it is explicitly designed to reduce the computational burden of conventional dynamical downscaling while preserving historical accuracy and spatiotemporal variability.

The framework uses conditional GANs in a three-step enhancement cascade. $v$3 performs a $v$4 spatial step from ERA5 at 30 km and hourly cadence to coarsened WTK at 10 km and hourly cadence. $v$5 performs a $v$6 spatial step from 10 km hourly to 2 km hourly. $v$7 performs a $v$8 temporal step from 2 km hourly to 2 km 5-minute output. The composite mapping is

$v$9

A modified content loss emphasizes extremes through spatial and temporal maxima and minima, and high-resolution topography is concatenated mid-network after spatial enhancement to improve flow realism over complex terrain.

The training corpus uses 2007–2009 and 2011–2013, with 2010 held out for validation. Each epoch comprises 100 batches of 64 samples drawn as spatiotemporal chunks of $w$0 low-resolution pixels. Reported training costs are 240 compute node hours for $w$1, 50 for $w$2, and 200 for $w$3, with 2,500, 7,000, and 10,000 epochs respectively. The first public product covers Ukraine, Moldova, and eastern Romania from January 2000 through December 2023 at 2-km spatial resolution and 5-minute temporal resolution, with winds at multiple hub heights spanning 10–200 m AGL.

Validation is reported against WFIP2, MADIS, and wind-farm observations. In aggregate CONUS WFIP2 statistics, Sup3rWind achieved MAE $w$4 m/s, MBE $w$5 m/s, Pearson $w$6, $w$7, MRQE $w$8, and KS $w$9, versus ERA5 values of MAE $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$0 m/s, MBE $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$1 m/s, $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$2, $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$3, MRQE $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$4, and KS $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$5. Over Ukraine MADIS stations, mean statistics were MAE $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$6 m/s, MBE $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$7 m/s, Pearson $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$8, $$G:\mathbb{R}^{H\times W\times Z\times C}\to \mathbb{R}^{H'\times W'\times Z'\times C},$$9, MRQE $C=3$0, and KS $C=3$1. Over Ukrainian wind farms, Sup3rWind reduced bias and improved MRQE and KS relative to ERA5, although ERA5 sometimes showed higher correlation in flatter terrain (Benton et al., 2024).

The system also includes ensemble-based uncertainty quantification through downscaling of 10-member ECMWF Ensemble of Data Assimilations inputs, interpolated from approximately 60 km and 3-hourly to 30 km and hourly before spatial downscaling. For the Ukraine product, one year of full spatiotemporal enhancement requires 40 compute node hours per year on 36-CPU nodes, distributed over roughly 30 nodes. Relative to a year-long 2-km, 5-minute WRF simulation for CONUS estimated at about 50,000 compute node hours, the GAN pipeline can produce a year in about 585 node hours on CPUs, yielding an $C=3$2 speedup, with GPU speedups reaching approximately $C=3$3 (Benton et al., 2024).

4. Superstatistical wind-field synthesis from sparse measurements

A third atmospheric usage treats Sup3rWind as a stochastic generator of synthetic wind fields with controlled intermittency (Friedrich et al., 2022). This model extends Gaussian inflow constructions such as the Mann model by representing the velocity field as a scale mixture of Gaussian fields with fluctuating covariances. For an $C=3$4-point velocity vector, the joint distribution is written as an integral over a mixing variable $C=3$5, so that intermittency enters through a hyper-distribution over covariance states rather than through a single Gaussian tensor.

The formulation is explicitly tied to the multipoint statistical description of turbulence and to K62-type scaling. Longitudinal structure functions satisfy

$C=3$6

which reduces to the K62 law for $C=3$7. In the study, fitting measured increment PDFs at several scales yields

$C=3$8

The covariance amplitude $C=3$9 is set from the measured root-mean-square velocity using

$H'=9H$0

The empirical basis is the GROWIAN campaign at Kaiser-Wilhelm-Koog, Germany, using two meteorological masts upstream of a 3 MW turbine. The dataset contains 16 propeller anemometers over a $H'=9H$1 area, sampled at 2.5 Hz, with 100 time series from a stable-stratification cluster selected for inference and validation. A 51.6 s excerpt is converted to a streamwise spatial coordinate via Taylor’s hypothesis. Synthetic fields are produced by Gaussian Fourier synthesis for each $H'=9H$2, followed by conditioning through a Gaussian bridge so that the reconstructed field hits all mast observations exactly.

The reported output is a reconstructed field on a $H'=9H$3 grid. The model matches measured correlation functions well over the resolved range, reproduces the increase of flatness as scale decreases, and captures the scale evolution and heavy tails of increment PDFs that a purely Gaussian baseline cannot reproduce. Because Fourier-space synthesis remains $H'=9H$4 and conditioning scales as $H'=9H$5 with $H'=9H$6 sensors, the approach remains orders of magnitude cheaper than LES or DNS for large three-dimensional domains (Friedrich et al., 2022).

5. Astrophysical superwinds, stellar winds, and relativistic outflows

In astrophysics, Sup3rWind is used for several physically distinct outflow classes. One major usage concerns starburst-driven superbubbles and galactic superwinds. Three-dimensional hydrodynamical simulations of starbursts report a blowout threshold tied to the analytic terminal speed $H'=9H$7: no blowout for $H'=9H$8, a transition regime for $H'=9H$9, and robust blowout insensitive to resolution and cooling once $W'=9W$0, with $W'=9W$1. Above threshold, simulated wind speed follows $W'=9W$2, with fitted $W'=9W$3 for T4 cooling and $W'=9W$4 for T1 cooling. Cooling to 10 K halves the mass fraction in warm H$W'=9W$5 gas while increasing the cold neutral and hot X-ray fractions, and the mass of cold gas transported into the lower halo does not depend on starburst strength. Complementary 3D superbubble calculations show that outflows appear by $W'=9W$6 Myr, reach heights of about 400 pc by $W'=9W$7 Myr, and that closely spaced clouds with $W'=9W$8 are hydrodynamically shielded strongly enough to keep $W'=9W$9 by $Z'=Z$0 Myr (Tanner et al., 2015, Suárez-López et al., 2024).

A second astrophysical cluster centers on radiative cooling, time-dependent ionization, and stellar-population evolution in superwinds. One-dimensional FLASH+MAIHEM models of starburst-driven superwinds incorporate radiative cooling and non-equilibrium ionization, showing strong departures from the adiabatic $Z'=Z$1 and $Z'=Z$2 expectations when cooling is efficient, and enhancing or radially shifting high-ion diagnostics such as C IV and O VI. Time-dependent feedback models driven by Starburst99 extend this by evolving $Z'=Z$3, $Z'=Z$4, and $Z'=Z$5, classifying solutions into adiabatic wind, adiabatic bubble, pressure-confined, catastrophic-cooling, momentum-conserving, and no-wind regimes through the temperature ratio $Z'=Z$6. Parameter-space surveys emphasize that high mass loading and reduced kinetic heating efficiency drive catastrophic cooling more strongly than metallicity alone, and that a hot superbubble does not necessarily imply an adiabatic outflow (Danehkar et al., 2023, Danehkar et al., 2024, Danehkar et al., 2021).

A third usage concerns super star cluster winds under strong cooling. In the semi-analytic framework of Wünsch et al., coupled to Starburst99 inputs, the wind can be quasi-adiabatic or bimodal. In the bimodal regime a stagnation radius $Z'=Z$7 forms, the inner region becomes thermally unstable, and mass accumulates as warm clumps while only the outer region drives a stationary wind. For a $Z'=Z$8, $Z'=Z$9 pc cluster, $\times 9$0 and $\times 9$1 produce bimodality from $\times 9$2 to $\times 9$3 Myr with $\times 9$4, whereas $\times 9$5 and $\times 9$6 yield bimodality from $\times 9$7 to $\times 9$8 Myr with $\times 9$9 (Wunsch et al., 2011).

AGN feedback provides yet another meaning. In moving-mesh AREPO simulations, small-scale ultra-fast AGN winds are injected through a spherical boundary with

$(i,j)$0

After a free-expansion phase, the wind thermalizes, inflates a hot shocked bubble, and powers energy-driven galactic super-winds. Above a critical luminosity, initially spherical small-scale winds generate bipolar outflows at $(i,j)$1 out to $(i,j)$2 kpc, suppressing star formation through rapid removal of dense nuclear gas and slower prevention of halo gas accretion. Momentum-driven solutions are described as rare because the Compton cooling radius is usually much smaller than the free-expansion radius (Costa et al., 2020).

The term also appears in analytic and compact-object contexts. Stationary, spherically symmetric galactic superwind solutions generalized to extended dark-matter and baryonic haloes show that external halo gravity can inhibit free superwinds and trigger earlier activation of open-box enrichment without changing the closed-box limit established in the inner-halo model (Añorve-Zeferino et al., 2010). Pc-resolution merger simulations with detailed stellar feedback produce multi-phase super-winds in all passages, with outflow rates up to $(i,j)$3, but with mass-loading efficiency that depends more on galaxy mass, size, and escape velocity than on merger state (Hopkins et al., 2013). In persistent supergiant X-ray binaries, Sup3rWind denotes ballistic line-driven supergiant winds whose large-scale morphology is fully determined by the reduced parameter set $(i,j)$4 and whose wind-fed accretion can form discs when the circularization radius exceeds the magnetospheric scale (Mellah et al., 2016). In pulsar-wind theory, the label is attached to a solution in which a striped wind converts into a strong superluminal electromagnetic wave precursor that decelerates, matches the external nebular pressure, and dissipates at the termination shock, thereby addressing the $(i,j)$5-problem (Mochol et al., 2013).

6. Conceptual relations, limitations, and nomenclature

The combined record suggests that Sup3rWind should not be read as the name of a single method, codebase, or dataset. In atmospheric machine learning alone it refers to at least three distinct constructs: a physics-informed SR-GAN for three-dimensional WRF winds, a renewable-energy reanalysis downscaling framework tied to ERA5 and WTK, and a superstatistical field generator for intermittent turbulence (Kurihana et al., 2023, Benton et al., 2024, Friedrich et al., 2022). The astrophysical usages are even more heterogeneous, spanning continuous stellar feedback, strong cooling in super star clusters, AGN wind injection, stationary halo-regulated superwinds, and compact-object wind problems (Wunsch et al., 2011, Costa et al., 2020, Añorve-Zeferino et al., 2010).

A common misconception is therefore to treat Sup3rWind as a single atmospheric product. The literature instead indicates a context-sensitive term whose meaning must be inferred from the governing equations, forcing mechanisms, and data sources of the specific study. Another misconception is that “high fidelity” in these works implies complete physical closure. The 3D atmospheric super-resolution model is explicitly preliminary, does not report explicit divergence-free enforcement, and does not fully specify training hyperparameters or the exact attention regularizer (Kurihana et al., 2023). The renewable-energy downscaling framework is trained on roughly six years of data, inherits limitations from the WTK target, and can show occasional noise in diurnal cycles or weaker correlation gains in flatter terrain (Benton et al., 2024). The superstatistical turbulence generator assumes local homogeneity and isotropy at small scales, contains no skewness in its present construction, and relies on Taylor’s hypothesis for space-time conversion (Friedrich et al., 2022).

Comparable caveats recur in astrophysical usages. Some studies are one-dimensional and spherically symmetric; others omit magnetic fields, cosmic rays, thermal conduction, or full radiative transfer; some focus on pre-supernova phases, while others neglect AGN feedback or multidimensional instability growth (Danehkar et al., 2024, Suárez-López et al., 2024, Añorve-Zeferino et al., 2010). Accordingly, the term is best interpreted not as a unified theory but as a label for multiscale wind modeling under strong computational or physical constraints. The literature suggests that its enduring commonality lies in that multiscale emphasis: resolving fine structure, preserving transport fidelity, or connecting small-scale driving to large-scale wind behavior.