Geko: A Multidisciplinary Research Overview

• Geko is a term with distinct definitions: GEKO for turbulence modeling in CFD, geko for JWST slitless spectroscopy, GECKO for digital pathology, and gecko for bio-inspired adhesion.
• In CFD, GEKO employs Bayesian optimization to calibrate tunable coefficients, enhancing predictions in separated flows and pressure recovery scenarios.
• Meanwhile, geko and GECKO implement Bayesian inference and contrastive pretraining for galaxy kinematics and whole-slide image analysis, while gecko adhesion studies model reversible attachment mechanics.

Geko denotes several distinct research objects in contemporary technical literature, and its meaning is determined almost entirely by capitalization and domain. In computational fluid dynamics, GEKO refers to the Generalized $k$-$\omega$ turbulence model and, more specifically in recent work, to the calibration of its closure coefficients for separated flows. In observational cosmology, $\texttt{geko}$ denotes the Grism Emission-line Kinematics tOol, a Bayesian forward-modeling package for JWST/NIRCam slitless spectroscopy. In computational pathology, GECKO denotes Gigapixel Vision-Concept Knowledge Contrastive pretraining for whole-slide images. A separate, non-acronymic literature on gecko adhesion addresses reversible biological attachment and detachment through cooperative elastic interactions (Kalia et al., 16 Feb 2025, Danhaive et al., 8 Oct 2025, Kapse et al., 1 Apr 2025, Puglisi et al., 2013).

1. Nomenclature and disciplinary scope

The term is not a single concept but a family of unrelated labels used in different research communities. The capitalization conventions are functionally significant.

Form	Expansion or use	Domain
GEKO	Generalized $k$-$\omega$ turbulence model	RANS turbulence modeling
$\texttt{geko}$	Grism Emission-line Kinematics tOol	JWST galaxy morpho-kinematics
GECKO	Gigapixel Vision-Concept Knowledge Contrastive pretraining	Histopathology machine learning
gecko	Biological organism used in adhesion mechanics studies	Biomechanics and bio-inspired adhesion

In the CFD literature, GEKO appears as a tunable two-equation turbulence model whose coefficients can be optimized for adverse-pressure-gradient and separated flows. In extragalactic astronomy, $\texttt{geko}$ is a Python package that forward-models slitless grism observations to infer intrinsic kinematic and morphological quantities. In digital pathology, GECKO is a contrastive pretraining framework for Multiple Instance Learning at whole-slide scale. The biological gecko literature, by contrast, studies reversible adhesion rather than an acronymic software or model name (Kalia et al., 16 Feb 2025, Danhaive et al., 8 Oct 2025, Kapse et al., 1 Apr 2025, Puglisi et al., 2013).

2. GEKO as a turbulence model

Within Reynolds-averaged Navier-Stokes modeling, GEKO is the Generalized $k$-$\omega$ model. The recent calibration literature emphasizes that it contains six tunable coefficients and can therefore be adapted to particular flow classes rather than used only with fixed defaults (Kalia et al., 16 Feb 2025).

The best-documented coefficients in the supplied literature are $C_\text{SEP}$ and $\omega$0. $\omega$1 controls boundary-layer separation by regulating eddy viscosity in adverse-pressure-gradient regions; higher $\omega$2 encourages separation, while lower values encourage attachment. $\omega$3 adjusts near-wall turbulence and wall-shear stress; increasing it boosts wall shear in non-equilibrium boundary layers. These controls make GEKO particularly relevant for converging-diverging channels, pressure recovery, separation, and reattachment, where conventional RANS closures are often challenged (Kalia et al., 16 Feb 2025).

The broader turbo-RANS literature treats GEKO as one of the turbulence models whose performance can be improved through systematic coefficient calibration. In that setting, GEKO is not altered at the closure-form level; rather, its predictive behavior is changed through optimized parameter values. This distinction matters because it preserves the existing model structure while shifting the calibration burden to a data-driven optimization loop (McConkey et al., 2023).

3. Bayesian optimization of GEKO coefficients

A major recent development is the use of turbo-RANS, a Bayesian-optimization framework for turbulence-model calibration. turbo-RANS introduces the Generalized Error and Default Coefficient Preference (GEDCP) objective function, designed to work with integral, sparse, or dense reference data while penalizing large departures from default coefficients. The framework is described as a semi-automated, open-source black-box calibration procedure that includes OpenFOAM, Ansys Fluent, STAR-CCM+, and solver-agnostic templates (McConkey et al., 2023).

For the converging-diverging channel case, GEKO was calibrated using sparse DNS data at $\omega$4. The training data consisted of wall coefficient of pressure $\omega$5 and coefficient of friction $\omega$6 at 10 selected points along the bottom wall. The workflow used a Python script to modify Ansys Fluent journal files, Sobol sequence sampling for initial exploration, and Bayesian optimization with a Gaussian Process surrogate for exploitation. The study reports initial bounds of $\omega$7 and $\omega$8, later expanding the lower bound of $\omega$9 to $\texttt{geko}$0. The calibration used a weighting factor $\texttt{geko}$1 to penalize coefficient excursions (Kalia et al., 16 Feb 2025).

The optimized coefficients were reported as

$\texttt{geko}$2

The optimized model improved wall quantity predictions, particularly reattachment, and improved streamwise velocity $\texttt{geko}$3 profiles at both $\texttt{geko}$4 and $\texttt{geko}$5. Pressure recovery $\texttt{geko}$6 was improved at both Reynolds numbers, whereas $\texttt{geko}$7 improved only modestly; the latter limitation was attributed to the inherent limitations of two-equation eddy-viscosity models in accurately resolving wall shear in adverse-pressure-gradient and separating flows. The same optimized coefficients generalized from DNS-referenced training at $\texttt{geko}$8 to LES comparisons at $\texttt{geko}$9, suggesting applicability across moderate Reynolds-number variation (Kalia et al., 16 Feb 2025).

4. $k$0 as a JWST/NIRCam morpho-kinematic inference tool

In astronomy, $k$1 is the Grism Emission-line Kinematics tOol, a Python package for modeling galaxy morphology and kinematics in JWST/NIRCam slitless spectroscopic observations. Its central problem is the morphology-kinematics degeneracy introduced by slitless dispersion, where spatial structure and velocity information are convolved along the dispersion axis. The package addresses this by forward-modeling the observation and performing Bayesian inference on the latent intrinsic parameters (Danhaive et al., 8 Oct 2025).

The default physical model combines a Sérsic surface-brightness profile with an arctangent rotation curve and a constant intrinsic velocity dispersion. The rotation field is written as

$k$2

where $k$3 is the asymptotic velocity and $k$4 is the turnover radius. The framework also uses the pressure-support-corrected circular velocity

$k$5

and dynamical mass

$k$6

Morphological priors can be supplied from imaging analyses such as Pysersic fits to JWST/NIRCam data (Danhaive et al., 27 Mar 2025).

Instrument modeling is explicit. $k$7 includes full PSF and LSF convolution, spectral and spatial pixelization, and projection from a synthetic 3D emission-line cube to the 2D detector plane. Inference is implemented within a Bayesian framework using $k$8 and $k$9, with HMC/NUTS described in the application paper and gradient-based sampling emphasized in the package paper. The tool recovers parameters such as effective radius, velocity dispersion, rotational velocity, rotational support, and dynamical mass, with typical run times of $\omega$0 minutes per galaxy on GPUs (Danhaive et al., 8 Oct 2025, Danhaive et al., 27 Mar 2025).

5. Astrophysical results enabled by $\omega$1

The first large statistical application reported a sample of 272 H$\omega$2 emitters in GOODS-S and GOODS-N at redshifts $\omega$3–$\omega$4, observed with JWST/NIRCam slitless spectroscopy and imaging from JADES, FRESCO and CONGRESS. Using $\omega$5, the study found $\omega$6 km/s and $\omega$7–$\omega$8 at $\omega$9–$\texttt{geko}$0, together with a slight increase in the fraction of rotationally supported systems from $\texttt{geko}$1 to $\texttt{geko}$2 from $\texttt{geko}$3 to $\texttt{geko}$4 for galaxies with masses $\texttt{geko}$5. The stated conclusion was that disks do not dominate the turbulent high-redshift galaxy population in the mass range probed (Danhaive et al., 27 Mar 2025).

A second application modeled the H$\texttt{geko}$6 morpho-kinematics of 163 galaxies at $\texttt{geko}$7–$\texttt{geko}$8 from FRESCO and CONGRESS, with JADES imaging. From the inferred rotational velocities and dispersions within $\texttt{geko}$9, the study derived dynamical, gas, baryonic, and dark-matter masses. It reported high median fractions of $\texttt{geko}$0 and $\texttt{geko}$1, with about two-thirds of systems being dark-matter dominated within $\texttt{geko}$2–$\texttt{geko}$3 kpc. Both $\texttt{geko}$4 and $\texttt{geko}$5 decrease with stellar mass, and $\texttt{geko}$6 shows a negative correlation with baryonic surface density $\texttt{geko}$7. The stellar Tully-Fisher relation showed a tentative offset to higher $\texttt{geko}$8 at fixed $\texttt{geko}$9 and substantial intrinsic scatter, which the study interpreted as evidence that the relation is only beginning to emerge at $k$0 (Danhaive et al., 16 Oct 2025).

A third study examined 213 galaxies at spectroscopic redshifts $k$1–$k$2 and used $k$3 to measure H$k$4 sizes and kinematics. At $k$5, it found average H$k$6 sizes larger than the stellar continuum, with $k$7 kpc and $k$8 kpc for galaxies with $k$9. The study reported no significant differences between stellar-continuum sizes at different wavelengths, suggesting that galaxies are not yet steadily growing inside-out at these epochs. It also found that $\omega$0 increases with distance above the star-forming main sequence, and that about half of elongated systems with $\omega$1 are not rotationally supported. The stated interpretation was that increased rotational support above the main sequence could be tracing accreting gas illuminated by H$\omega$2 emission (Danhaive et al., 7 Oct 2025).

6. GECKO in computational histopathology

In digital pathology, GECKO stands for Gigapixel Vision-Concept Knowledge Contrastive pretraining. The method addresses a different problem from both GEKO and $\omega$3: pretraining a Multiple Instance Learning aggregator so that whole-slide image embeddings can be derived from patch-level representations without supervision. Its motivation is that prior multimodal MIL approaches often require auxiliary modalities such as transcriptomics, which increases cost and limits scalability when paired modalities are unavailable (Kapse et al., 1 Apr 2025).

GECKO constructs a concept prior by computing similarity between each whole-slide-image patch and textual descriptions of predefined pathology concepts. It then uses a dual-branch MIL network: one branch aggregates patch embeddings into a WSI-level deep embedding, and the other aggregates the concept prior into a WSI-level concept embedding. The two embeddings are aligned with a contrastive objective. When auxiliary modalities such as transcriptomics are available, GECKO can incorporate them without changing the basic framework (Kapse et al., 1 Apr 2025).

The reported empirical claim is that, across five diverse tasks, GECKO consistently outperformed prior unimodal and multimodal pretraining approaches while also delivering clinically meaningful interpretability. Its interpretability is tied to the concept embedding itself, which is intended to bridge computational predictions and pathology expertise at slide level rather than only through attention heatmaps (Kapse et al., 1 Apr 2025).

7. Gecko adhesion mechanics and the non-acronymic literature

Separate from the acronymic usages, the biological gecko literature studies reversible adhesion. A mechanically explicit account was given through a model in which the detachment force depends not only on the properties of individual adhesive units but also on the elastic interaction among them. The model implies active muscle control and distinguishes two detachment modes: delocalized pull-off and localized peeling (Puglisi et al., 2013).

The key parameter is the coupling stiffness $\omega$4, which sets the scale of cooperative interaction among adhesive units. In the continuum description, the maximum detachable force is

$\omega$5

For stiff pads with strong coupling, detachment is localized and peeling-like, with

$\omega$6

For soft pads with weak coupling, decohesion is spread out and pull-off-like, with

$\omega$7

The same study argued that active variation of pad stiffness can reduce the detachment threshold by over an order of magnitude via a two-orders-of-magnitude change in $\omega$8 (Puglisi et al., 2013).

The work also proposed a hierarchical interpretation in which the same mechanism operates from spatula to seta to lamella to pad to foot, with adhesion force scaling as $\omega$9 and $C_\text{SEP}$0. This suggests a design principle for artificial adhesives: controllable cooperative stiffness can switch a system between strong attachment and easy release without changing the local adhesive chemistry (Puglisi et al., 2013).