SWIRL: Vortex, Workflow, and RL Systems

• SWIRL is an acronym referring to three distinct systems: one for vortex identification in fluid dynamics, one as a workflow intermediate representation, and one for staged reinforcement learning.
• The vortex identification method uses the Rortex criterion and clustering of estimated vortex centers to reliably detect coherent swirls in noisy dynamical flows.
• The workflow IR and staged RL variants employ formal rewriting rules and interleaved optimization to enhance distributed computation efficiency and multi-agent training stability.

In recent arXiv literature, SWIRL denotes several technically distinct research artifacts rather than a single unified method. The name is used for “SWirl Identification by Rotation-centers Localization,” an automated vortex-identification algorithm for fluid and solar-atmosphere flows (Cuissa et al., 2022); “Scientific Workflow Intermediate Representation Language,” a low-level intermediate representation for distributed scientific workflow execution plans (Colonnelli et al., 2024); and “Staged Workflow for Interleaved Reinforcement Learning,” a multi-agent reinforcement-learning workflow developed for mobile GUI control and related tasks (Lu et al., 27 Aug 2025). The recurrence of the acronym across solar physics, workflow systems, and reinforcement learning suggests a shared emphasis on structured intermediate representations, but the underlying semantics, objectives, and validation regimes are unrelated.

1. Acronym, scope, and disciplinary usage

Expansion	Domain	Stated role
SWirl Identification by Rotation-centers Localization	Fluid dynamics and solar physics	Automated vortex identification (Cuissa et al., 2022)
Scientific Workflow Intermediate Representation Language	Scientific workflow systems	Low-level compilation target for workflow execution plans (Colonnelli et al., 2024)
Staged Workflow for Interleaved Reinforcement Learning	Multi-agent RL and LVLM-based agents	Sequence of single-agent RL tasks for multi-agent systems (Lu et al., 27 Aug 2025)

The earliest SWIRL in the supplied corpus is the vortex-identification algorithm introduced to address the absence of a universally accepted vortex definition and the difficulty of identifying coherent vortices in turbulent, noisy, and highly dynamical flows (Cuissa et al., 2022). A later usage reassigns the acronym to a workflow IR designed for interoperability, distributed-by-design execution, and optimization of data movement in scientific computing (Colonnelli et al., 2024). A third usage appears in multi-agent reinforcement learning, where SWIRL denotes a staged training workflow that reformulates MARL into alternating single-agent optimization blocks (Lu et al., 27 Aug 2025).

This multiplicity is not merely terminological. The three systems operate on different primitives: velocity gradients and curvature centers in one case, location-aware execution traces in another, and policy-update stages in the third. Any technical discussion of SWIRL therefore requires explicit domain disambiguation.

2. SWIRL as vortex identification by rotation-centers localization

The SWIRL algorithm was introduced as a hybrid method that combines “the rigor of mathematical criteria with the global perspective of morphological techniques” and aims to avoid “the use of a threshold that could potentially prevent the detection of weak vortices” (Cuissa et al., 2022). Its core local quantity is the Rortex criterion, adopted because it was argued to measure “the rigid-body rotational part of the flow alone” and to be less biased by pure or intrinsic shears than vorticity or swirling strength. In the formulation used by the method,

$$R = \boldsymbol{\omega}\cdot\boldsymbol{u}_{\rm r} - \sqrt{(\boldsymbol{\omega}\cdot\boldsymbol{u}_{\rm r})^2 - \lambda^2},$$

where $\boldsymbol{\omega}$ is vorticity, $\boldsymbol{u}_{\rm r}$ is the normalized real eigenvector of the velocity-gradient tensor, and $\lambda$ is the swirling-strength criterion (Cuissa et al., 2022).

The distinctive step in SWIRL is the conversion of local rotational information into Estimated Vortex Centers (EVCs). For each grid cell with local curvature, the method estimates a center of rotation from the local radius of curvature and radial direction. Vortex identification is then recast as a clustering problem: grid cells that belong to the same coherent vortex should produce EVCs clustered around a common core. The algorithm maps EVCs to the grid, forms a G-EVC map through signed grid cardinalities, separates clockwise and counterclockwise populations, and applies a modified CFSFDP density-peaks clustering procedure to identify candidate vortices (Cuissa et al., 2022).

This design explicitly distinguishes coherent vortices from mere local curvature patches. In the paper’s test program, SWIRL was validated on Lamb-Oseen vortices, on noisy two-vortex configurations, and on a 2D MHD Orszag-Tang case, where it identified 37 swirls (Cuissa et al., 2022). The method was implemented in open-source Python, and the authors reported that clustering dominated runtime. They also noted several limitations: the implementation was 2D, center estimation was “not strictly Galilean invariant,” and identification remained sensitive to parameters such as stencil count and clustering-cleaning settings (Cuissa et al., 2022).

3. Solar-atmosphere deployment and relation to solar swirls

A subsequent application paper assessed SWIRL on realistic three-dimensional radiative MHD simulations of the solar atmosphere produced with CO5BOLD (Cuissa et al., 2023). In that study, SWIRL was applied to 30 time instances spanning 2 h of physical time. The identified structures were described as “spiraling instantaneous streamlines of the horizontal component of the flow,” and part of the identified swirls formed “three-dimensional coherent structures that are generally rooted in magnetically dominated intergranular lanes and extend vertically into the chromospheric layers” (Cuissa et al., 2023).

The statistical results reported for this deployment were substantially more aggressive in scale than earlier manually or threshold-based solar-swirl studies. The average number densities were given as 1 Mm$^{-2}$ in the photosphere and 4 Mm$^{-2}$ in the chromosphere, while the average radius remained 50–60 km throughout the simulated atmosphere (Cuissa et al., 2023). The same paper also reported “an approximately linear correlation between the rotational speed of chromospheric swirls and the local Alfvén speed,” with fitted power-law exponent $b = 1.16$ and Pearson coefficient $r_P = 0.45$ (Cuissa et al., 2023). In a separate Alfvénic interpretation of chromospheric swirls, simulations using CO$^5$BOLD argued that upward-propagating swirls behave as torsional Alfvén pulses, with strong correlation between plasma swirling strength and magnetic swirling strength (Battaglia et al., 2021).

The SWIRL algorithm sits within a broader solar-swirl literature that uses different operational definitions. In observational chromospheric studies, chromospheric swirls are defined as the observational signatures of rotating magnetic field structures, also called magnetic tornadoes, and are identified from dark ring, ring-fragment, or spiral features in the Ca II 854.2 nm line core that rotate, persist for several minutes, show vertical velocities of at least about $2\ \mathrm{km\,s^{-1}}$, and coincide with a photospheric magnetic concentration (Wedemeyer et al., 2013). Automated detection in H$\boldsymbol{\omega}$0 -0.2 Å observations found a mean of $\boldsymbol{\omega}$1 swirls in the field of view at any given time, a mean surface density of about 0.08 swirls Mm$\boldsymbol{\omega}$2, and a survival-analysis mean lifetime of $\boldsymbol{\omega}$3 min (Dakanalis et al., 2022). In the photosphere, ASDA-based studies defined swirls from horizontal velocity fields using the Graftieaux criteria and reported average radii of about 290–307 km, together with strong sensitivity to spatial resolution (Liu et al., 2018, Xie et al., 2024).

These comparisons indicate that SWIRL, ASDA, and chromospheric morphology-based classifiers do not measure identical objects. SWIRL detects coherent vortical structures in velocity fields by clustering estimated rotation centers; ASDA detects swirls in inferred horizontal flows through $\boldsymbol{\omega}$4 and $\boldsymbol{\omega}$5 vortex indicators; chromospheric observational definitions rely on line-core morphology, Doppler behavior, and magnetic footpoint co-location (Liu et al., 2018, Wedemeyer et al., 2013).

4. SWIRL as a scientific workflow intermediate representation language

In scientific computing, SWIRL was later redefined as Scientific Workflow Intermediate Representation Language, a “low-level, formal intermediate representation for scientific workflow execution plans” (Colonnelli et al., 2024). The paper was explicit that this SWIRL is “not designed for human interaction” and is instead intended to serve as a low-level compilation target for heterogeneous workflow systems. The stated goals were interoperability across workflow systems, distributed-by-design execution, automatic optimization of data movement, and formal guarantees for translation and optimization (Colonnelli et al., 2024).

The language formalizes a workflow as a distributed system of location configurations. Its operational primitives are exec, send, and recv, combined by sequential and parallel composition. The syntax reported in the paper includes workflow systems of the form

$\boldsymbol{\omega}$6

with execution traces

$\boldsymbol{\omega}$7

and primitive actions

$\boldsymbol{\omega}$8

Communications do not consume data at the sender; instead, the data element is copied to the receiver (Colonnelli et al., 2024).

A central technical claim of this SWIRL is that the semantics supports a correctness framework based on structural congruence, reduction rules, and a Church–Rosser-style concurrency property. The paper further introduced rewriting rules to eliminate redundant local communications and repeated transmissions of the same data item, and it justified these optimizations through weak barbed bisimulation, establishing operational correspondence between original and optimized workflow systems (Colonnelli et al., 2024).

The reference toolchain was developed with ANTLR Python3 bindings. The implementation included a grammar, a translation framework, concrete translators for languages such as CWL, DAX (Pegasus), and GWF (Galaxy Workflow Format), an optimizer, and a compiler capable of generating multithreaded Python bundles per location with TCP-socket communication (Colonnelli et al., 2024). A simplified 1000 Genomes workflow served as the case study for the encoding.

5. SWIRL as a staged workflow for interleaved reinforcement learning

A third technical meaning of SWIRL appears in reinforcement learning: Staged Workflow for Interleaved Reinforcement Learning (Lu et al., 27 Aug 2025). The core claim of this framework is that MARL can be made more practical and stable by “turning a coupled multi-agent optimization problem into a sequence of single-agent RL problems.” At any point, one agent is optimized while the others are frozen, so the current update can be treated with standard single-agent RL machinery (Lu et al., 27 Aug 2025).

The framework is organized into two stages. Stage 1 is a warm-up initialization phase in which agents learn their roles separately. In the mobile-GUI instantiation, the Navigator is initialized with lightweight Chain-of-Thought supervised fine-tuning on instruction–action pairs, while the Interactor is bootstrapped with RL using fixed planner outputs (Lu et al., 27 Aug 2025). Stage 2 is a round-level interleaved RL procedure: one agent is chosen, its counterpart is frozen, the chosen agent is updated for several inner steps, and then the roles are swapped. The paper described the rolling baseline policy as

$\boldsymbol{\omega}$9

and stated a stepwise safety bound, a cross-round monotonic improvement theorem, and convergence guarantees on return (Lu et al., 27 Aug 2025).

For mobile GUI control, SWIRL instantiates a two-agent system in which the Navigator converts language, history, and screen context into a structured low-level instruction, and the Interactor grounds that instruction into an executable atomic GUI action (Lu et al., 27 Aug 2025). The reported training setup used Qwen2.5-VL-3B for both agents, 1,500 samples in Stage 1, 2,000 additional samples in Stage 2, and a total of 3,500 examples. Evaluation was zero-shot on AndroidControl-High, GUIOdyssey, AndroidControl-Low, GUI-Act-Web, OmniAct-Web, and OmniAct-Desktop (Lu et al., 27 Aug 2025).

The paper reported that the dual-agent SWIRL system with two 3B models outperformed OS-Atlas-7B by 18.8 points overall and GUI-R1-7B by 7.55 points overall on the high-level GUI setting (Lu et al., 27 Aug 2025). The same work also applied SWIRL to a two-agent mathematics system with a Teacher and Student, using Qwen2.5-Coder-3B-Instruct, and reported improvements over MARFT on MATH500, CMATH, and GSM8K (Lu et al., 27 Aug 2025). Within this meaning of the acronym, SWIRL is therefore a training protocol rather than a detector or a language.

6. Comparative interpretation and conceptual distinctions

The three main SWIRL systems share almost no operational substrate. The vortex-identification SWIRL extracts coherent rotational structures from velocity fields by estimating rotation centers and clustering them (Cuissa et al., 2022). The workflow-language SWIRL represents distributed scientific workflows as location-specific traces of exec, send, and recv actions (Colonnelli et al., 2024). The reinforcement-learning SWIRL organizes optimization into alternating stages in which one agent is updated while the others remain fixed (Lu et al., 27 Aug 2025).

This suggests that the commonality of the name lies less in method than in architecture. Each SWIRL introduces an intermediate object that mediates between raw input and higher-level inference: EVC clusters between velocity gradients and vortex identities, workflow IR traces between user-facing workflow descriptions and execution bundles, and staged single-agent subproblems between coupled MARL objectives and trainable policies. Even so, the criteria of success are domain-specific. For the vortex-identification algorithm, the relevant issues are coherent structure, shear contamination, and parameter sensitivity (Cuissa et al., 2022). For the workflow IR, they are operational correspondence, communication optimization, and compilation across heterogeneous systems (Colonnelli et al., 2024). For the RL workflow, they are training stability, monotonic improvement, memory efficiency, and zero-shot benchmark performance (Lu et al., 27 Aug 2025).

For technical usage, “SWIRL” therefore functions as a domain-dependent proper name. In solar and fluid contexts it ordinarily refers to SWirl Identification by Rotation-centers Localization; in scientific workflow systems it refers to Scientific Workflow Intermediate Representation Language; and in LVLM-based agent training it refers to Staged Workflow for Interleaved Reinforcement Learning (Cuissa et al., 2022, Colonnelli et al., 2024, Lu et al., 27 Aug 2025).