DragonFly: Multi-Domain Research Overview

• DragonFly is a multi-disciplinary designation referring to diverse instruments, algorithms, models, and missions across astronomy, machine learning, optimization, planetary exploration, and bio-inspired biomechanics.
• The Dragonfly Telephoto Array and its spectral mapper achieve ultra–low surface brightness and narrowband imaging with innovative optical designs that minimize scattered light and enhance sensitivity.
• In machine learning and HPC networking, DragonFly drives multi-resolution visual encoders, swarm-intelligence algorithms, and high-radix interconnects, enabling fine-grained sensing and optimized performance in complex environments.

DragonFly, more commonly also written as Dragonfly, is a recurrent designation in contemporary research rather than a single object. In the literature it denotes, among other things, a refractive astronomical array for ultra–low surface brightness imaging, an ultranarrow-band spectral-line mapper derived from that array, a large multimodal model for fine-grained vision-language understanding, a swarm-intelligence optimization method and associated software libraries, a Titan rotorcraft mission context including navigation and atmospheric-tide studies, a high-radix interconnect topology and its associated performance-analysis models, a mmWave backscatter localization system, a pulsar wind nebula, and several dragonfly-inspired studies in biomechanics, migration, and micro-ornithopter dynamics (Abraham et al., 2014, Thapa et al., 2024, Rahman et al., 2020, Schilling et al., 2023).

1. Nomenclature and research scope

The term appears across multiple disciplines with distinct technical meanings. In astronomy, Dragonfly refers to the Dragonfly Telephoto Array and its narrowband descendant, the Dragonfly Spectral Line Mapper. In machine learning, DragonFly names a large multimodal model, a deep reinforcement learning library, and a Bayesian optimization package. In optimization theory, Dragonfly Algorithm denotes a swarm-intelligence metaheuristic. In planetary exploration, Dragonfly denotes the Titan rotorcraft mission context, including navigation filtering and atmospheric-tide science. In high-performance computing, Dragonfly denotes a low-diameter interconnect topology; related work studies workload interference and surrogate runtime prediction. In sensing, DragonFly denotes a single-radar mmWave backscatter localization system. In astrophysics and biology, Dragonfly refers respectively to the pulsar wind nebula G75.2+0.1 and to empirical and bio-inspired studies of dragonfly flight and migration (Abraham et al., 2014, Thapa et al., 2024, Viquerat et al., 30 Apr 2025, Kandasamy et al., 2019, Kang et al., 2024, Harisha et al., 7 Jul 2025, Woo et al., 2023, Ranjan et al., 2022).

Domain	Referent	Core description
Astronomy	Dragonfly Telephoto Array	Robotic refracting array for ultra low surface brightness imaging
Astronomy	Dragonfly Spectral Line Mapper	120-lens ultranarrow-band spectral-line mapper
Multimodal ML	DragonFly	Multi-resolution zoom-in vision-LLM
Optimization	Dragonfly Algorithm	Swarm-intelligence metaheuristic with five social behaviors
RL software	Dragonfly	JSON-driven modular deep reinforcement learning library
Bayesian optimization	Dragonfly	Scalable and robust Gaussian-process BO library
Planetary exploration	Dragonfly mission context	Titan navigation filter and atmospheric-tide probe concept
HPC networking	Dragonfly interconnect	High-radix, low-diameter network topology
Radar sensing	DragonFly	Single mmWave radar 3D localization of dynamic backscatter tags

A plausible implication is that the persistence of the name reflects a shared emphasis on mobility, distributed structure, or fine-grained sensing, but the underlying systems are otherwise unrelated.

2. Astronomical instrumentation and ultra–low surface brightness imaging

The Dragonfly Telephoto Array is a robotic imaging system optimized for detecting extended ultra low surface brightness structures in the nearby Universe. Its original implementation used eight Canon 400 mm f/2.8 telephoto lenses coupled to eight science-grade CCD cameras, yielding an imaging capability equivalent to a 0.4 m aperture f/1.0 refractor with a 2.6 deg × 1.9 deg field of view. The design eliminates central obstruction, reflective surfaces, and much of the large-radius scattered light characteristic of reflecting telescopes; measured stellar halos show a factor ≈ 5–10 less scattered light at radii $> 5'$ than the well-baffled Burrell Schmidt. The array reaches $\mu_B < 30$ mag arcsec$^{-2}$ in $\approx 10$ h without binning or foreground-star removal, and $\mu_B < 32$ mag arcsec$^{-2}$ in 50–100 h with modest binning and foreground-star masking/removal. For $N$ identical lenses of focal length $f$ and diameter $D$, the effective aperture and f-number are

$D_{\mathrm{eff}} = \sqrt{N}\,D,$

$\mu_B < 30$0

which for $\mu_B < 30$1 and $\mu_B < 30$2 yields an effective $\mu_B < 30$3 system (Abraham et al., 2014).

The same optical philosophy was extended to ultranarrow-band imaging in the Dragonfly Filter-Tilter concept. By placing interference filters at the entrance pupil rather than in a converging beam, Dragonfly preserves a narrow, stable passband and avoids angle-induced broadening. The 2020 narrowband design targeted $\mu_B < 30$4 nm, with a $\mu_B < 30$5 nm tuning range produced by tilting the filter. The central wavelength follows

$\mu_B < 30$6

allowing the same filter to be tuned across nearby extragalactic recession velocities while retaining high contrast for diffuse line emission (Lokhorst et al., 2020).

This concept matured into the Dragonfly Spectral Line Mapper, a 120-lens mosaic telescope built in four phases from March 2022 to November 2023. Each lens is a Canon 400 mm f/2.8 IS II with an SBIG Aluma CCD694 camera, giving a plate scale of $\mu_B < 30$7 arcsec pixel$\mu_B < 30$8 and a field of view of $\mu_B < 30$9. The array comprises four mounts of 30 lenses each, with effective collecting area $^{-2}$0 m$^{-2}$1. Ultra-narrow $^{-2}$2 nm filters target H$^{-2}$3, [N II], and [O III], and the Filter-Tilter provides $^{-2}$4 tuning, corresponding to velocity coverage of $^{-2}$5–3900 km s$^{-2}$6 for H$^{-2}$7 and [O III]. The system is designed to reach H$^{-2}$8 surface-brightness sensitivity of $^{-2}$9 erg s$\approx 10$0 cm$\approx 10$1 arcsec$\approx 10$2 over a $\approx 10$3 field with velocity resolution $\approx 10$4 km s$\approx 10$5 (Chen et al., 2024).

The three-lens pathfinder demonstrated the method on sky. It used 152 mm clear-aperture, 3 nm interference filters mounted in Filter-Tilters on three Canon 400 mm f/2.8 lenses, with $\approx 10$6 s narrowband exposures and matching tilt- and pointing-specific flats. In the M81/M82 field it achieved an H$\approx 10$7 surface-brightness limit of $\approx 10$8 erg cm$\approx 10$9 s$\mu_B < 32$0 arcsec$\mu_B < 32$1 at $\mu_B < 32$2 on $\mu_B < 32$3 scales, and discovered a giant shell-like cloud of ionized gas near M82 with angular extent $\mu_B < 32$4, projected physical extent $\mu_B < 32$5 kpc, projected distance $\mu_B < 32$6 kpc from M82, average H$\mu_B < 32$7 surface brightness $\mu_B < 32$8 erg cm$\mu_B < 32$9 s$^{-2}$0 arcsec$^{-2}$1, and follow-up line ratios $^{-2}$2 and $^{-2}$3 (Lokhorst et al., 2022, Lokhorst et al., 2022).

3. Machine learning, optimization, and software systems

In multimodal machine learning, DragonFly is a large multimodal model designed for fine-grained visual understanding and region-level reasoning. Its core mechanism is a multi-resolution encoding pipeline combined with zoom-in sub-crop selection. A CLIP ViT-B/32 image encoder processes one low-resolution crop at $^{-2}$4, four medium-resolution crops chosen from $^{-2}$5, and 24 high-resolution crops chosen from $^{-2}$6; a region-selection step then keeps the top-3 high-resolution crops for each medium-resolution region, yielding $^{-2}$7 crops and approximately $^{-2}$8 visual tokens. Mean-pooled projected tokens define selection summaries,

$^{-2}$9

and the selected visual tokens are concatenated directly with text tokens for Llama 3 8B-chat. On the final model, reported scores include AI2D $N$0, ScienceQA $N$1, ChartQA $N$2, MMMU-val $N$3, MMVet $N$4, POPE $N$5, and COCO CIDEr $N$6. The biomedical variant DragonFly-Med reports SLAKE accuracy $N$7, Path-VQA accuracy $N$8, and strong captioning results on IU X-Ray, Peir Gross, and ROCO (Thapa et al., 2024).

Dragonfly Algorithm is a swarm-intelligence metaheuristic that imitates static and dynamic dragonfly swarming. It uses five behavioral vectors—separation, alignment, cohesion, attraction toward food, and distraction from enemies—and updates a step vector $N$9 and a position $f$0 through

$f$1

$f$2

When no neighbors exist, a random walk such as Lévy flight is used. The surveys distinguish continuous DA, Binary DA, and multi-objective DA, and discuss variants including MHDA, INMDA, HAD, Brownian-DA, chaotic DA, and MODA/RMODA. The surveyed literature describes strong performance on many small- to medium-scale engineering problems but weaker results on large-scale CEC-C2019 benchmarks, where GWO outperformed DA overall and DA was slowest in runtime among compared algorithms (Rahman et al., 2020, Rahman et al., 2019).

Dragonfly also names two software libraries. One is a modular deep reinforcement learning library centered on JSON-driven experiments, a generic factory pattern, and TensorFlow 2. It supports A2C, PPO, DQN, DDPG, TD3, and SAC; uses types.SimpleNamespace for configuration; and includes CPU-oriented features such as parallel workers, ring buffers, a bootstrapping strategy for on-policy scaling, PCA and auto-encoder state representation learning, and a separable-environment trainer. On the Beacon shkadov-v0 environment, the bootstrapping strategy preserves PPO performance up to 16 parallel environments and only slightly degrades up to 32 (Viquerat et al., 30 Apr 2025). The other is a Bayesian optimization library for expensive black-box functions. It combines Gaussian-process surrogates, acquisition portfolios, hyperparameter portfolios, Add-GP-UCB for high-dimensional additive structure, BOCA-style multi-fidelity selection, mixed-variable kernels, constraints via Python predicates, and evolutionary acquisition optimization over non-Euclidean domains. Its GP posterior is the standard

$f$3

$f$4

and the package targets hyperparameter tuning, multi-fidelity model selection, and neural architecture search (Kandasamy et al., 2019).

4. Titan exploration, navigation, and atmospheric-tide science

In planetary exploration, Dragonfly denotes the Titan rotorcraft mission context. The preliminary design of the Dragonfly Mobility subsystem’s navigation filter uses an Extended Kalman Filter running at $f$5 Hz to estimate errors in a high-rate inertial navigator propagated at $f$6 Hz from IMU data sampled at $f$7 Hz. Two Navigators run in parallel, one per IMU, though only the primary state is fused. The EKF state has 47 states, including position, velocity, multiplicative attitude error, range to current image, perpendicular ground distance, accelerometer and gyro biases for both IMUs, pressure-sensor biases, ETS velocimetry bias, atmospheric scale-height error, ground surface normal, global heading errors, and multiple augmented positions tied to current and breadcrumb images. Measurements come from pressure altimetry, lidar altimetry, and visual odometry from the Navigation Camera through the Electro-optical Terrain Sensing module. Key design elements for multi-kilometer flights include two reference images for velocimetry, explicit modeling of correlated image errors through a 2D FOGM bias $f$8, and a selective approximation to SLAM using only five augmented position vectors instead of a full landmark map. Simulations reported heading initialization to $f$9 ($D$0) after about 60 minutes of gyrocompassing, breadcrumb-relative lateral errors $D$1 m after historic-breadcrumb processing, and $D$2 reduction in vertical-velocity uncertainty during terminal descent when lidar partials are included above $D$3–60 m (Schilling et al., 2023).

A separate Titan study treats Dragonfly as a pressure-observing platform for probing Titan’s interior through gravitational atmospheric tides. Saturn’s eccentricity tide produces a degree-2 surface pressure signal whose dynamically relevant forcing is modulated by the complex Love-number combination $D$4. The surface-pressure response is

$D$5

with predicted interior-controlled parameters

$D$6

Because Titan’s interior deformation strongly weakens the direct gravitational forcing, the residual surface-pressure amplitude is only $D$7 Pa, with phase shift $D$8–20 hours, and the lower-tropospheric tidal winds are of order $D$9 m s$D_{\mathrm{eff}} = \sqrt{N}\,D,$0. The study argues that mission-long Dragonfly pressure measurements could estimate the real and imaginary parts of $D_{\mathrm{eff}} = \sqrt{N}\,D,$1 with precision $D_{\mathrm{eff}} = \sqrt{N}\,D,$2–0.03, thereby constraining Titan’s ice-shell thickness, surface heat flux, and ocean density (Charnay et al., 2021).

5. Networks, workload prediction, and radar localization

In high-performance computing, Dragonfly is a high-radix, low-diameter interconnect that organizes routers into groups with local intra-group links and global inter-group links. In the 1,056-node system studied for routing interference, the canonical parameters are $D_{\mathrm{eff}} = \sqrt{N}\,D,$3 routers per group, $D_{\mathrm{eff}} = \sqrt{N}\,D,$4 global links per router, $D_{\mathrm{eff}} = \sqrt{N}\,D,$5 endpoints per router, $D_{\mathrm{eff}} = \sqrt{N}\,D,$6 groups, and router radix $D_{\mathrm{eff}} = \sqrt{N}\,D,$7. End-to-end communication completes in three hops: local, global, local. The study compares UGALg, UGALn, PAR, and a reinforcement-learning-based Q-adaptive routing policy. Using nine workloads with different communication patterns, it introduces message injection rate and peak ingress volume as two intensity metrics and shows that Q-adaptive substantially reduces interference. In mixed six-application workloads, adaptive routing increased communication time by $D_{\mathrm{eff}} = \sqrt{N}\,D,$8 on average relative to interference-free baselines, while Q-adaptive cut this added interference by $D_{\mathrm{eff}} = \sqrt{N}\,D,$9 on average, reduced average packet latency to $\mu_B < 30$00s, reduced the $\mu_B < 30$01th percentile latency to $\mu_B < 30$02s, and improved aggregated throughput from $\mu_B < 30$03 to $\mu_B < 30$04 GB ms$\mu_B < 30$05 (Kang et al., 2024).

Runtime prediction on Dragonfly systems is addressed by SMART, a surrogate model combining a graph convolutional encoder, a temporal Transformer, and a Time-LLM module. The graph is defined over router ports, with edges representing within-router cliques and physical links, and telemetry snapshots are aggregated every $\mu_B < 30$06s. The GCN uses

$\mu_B < 30$07

and the model forecasts per-iteration runtimes for active terminal ports. On simulated 1,056-node Dragonfly datasets, SMART reports consistent MAPE reductions over LAST, MEAN, LSTM, and DCRNN; examples include D1-Cont LAMMPS $\mu_B < 30$08 versus $\mu_B < 30$09 for LSTM and $\mu_B < 30$10 for DCRNN, and D2-Rand NN $\mu_B < 30$11 versus $\mu_B < 30$12 and $\mu_B < 30$13. Average inference time is $\mu_B < 30$14 s, versus PDES per-iteration times of $\mu_B < 30$15–$\mu_B < 30$16 s (Wang et al., 14 Nov 2025).

DragonFly also names a single-anchor mmWave backscatter localization system for GPS-denied environments. It uses a single cross-polarized MIMO FMCW radar with 2 Tx and 8 Rx channels at 24 GHz, together with backscatter tags that include a biconvex PTFE dielectric lens, a 29-element cross-polarized patch array, ultra-low-power switches, and a baseband oscillator. The tag power budget is $\mu_B < 30$17W at 250 kHz modulation. Localization is based on intra-chirp frequency coding, range/AoA FFT processing, and a Doppler-disambiguation method for elevation under time-division multiplexed Tx. The basic relations are

$\mu_B < 30$18

and the radar backscatter link budget follows

$\mu_B < 30$19

The reported system tracks multiple highly dynamic tags with median 3D accuracy $\mu_B < 30$20 cm, at speeds on the order of $\mu_B < 30$21 m s$\mu_B < 30$22, accelerations on the order of $\mu_B < 30$23 m s$\mu_B < 30$24, and ranges up to $\mu_B < 30$25 m (Harisha et al., 7 Jul 2025).

6. Astrophysical, biological, and bio-inspired referents

In high-energy astrophysics, Dragonfly is the common name of the pulsar wind nebula G75.2+0.1 powered by PSR J2021+3651. It is spatially associated with LHAASO J2018+3651, detected up to a maximum photon energy of $\mu_B < 30$26 PeV. NuSTAR resolves a hard X-ray inner nebula of radius $\mu_B < 30$27 with a power-law spectrum of $\mu_B < 30$28 and energy-dependent shrinkage from $\mu_B < 30$29 in 3–6 keV to $\mu_B < 30$30 in 6–20 keV, implying synchrotron burn-off and $\mu_B < 30$31G at 3.5 kpc. XMM-Newton reveals an outer nebula extending $\mu_B < 30$32 west with $\mu_B < 30$33. Multiwavelength modeling with VLA, XMM, and HAWC data supports a leptonic PeVatron interpretation with maximum injected particle energy $\mu_B < 30$34 PeV and $\mu_B < 30$35G (Woo et al., 2023).

In biomechanics, dragonflies are studied as four-wing flyers with strongly spanwise-dependent aerodynamics. Time-resolved stereo PIV on Pantala flavescens in forward flight with hindwing-leading phasing $\mu_B < 30$36 showed that the forewing develops a coherent leading-edge vortex that forms without hindwing influence, while the hindwing exhibits TEV-dominated root flow, LEV enhancement in the inner span when fore- and hindwings move oppositely in close proximity, LEV suppression at mid-span by forewing downwash, and wake-capture-driven LEV development in the outer span near stroke end. This suggests different spanwise roles in force generation. A separate modeling study of a dragonfly-inspired four-wing micro ornithopter formulates the full dynamics on $\mu_B < 30$37, with four rigid wings attached by spherical joints and quasi-steady blade-element aerodynamics. Its reduced body equations take the form

$\mu_B < 30$38

$\mu_B < 30$39

and simulation results emphasize the importance of inertial wing–body interaction terms $\mu_B < 30$40 and $\mu_B < 30$41 in near-hover (Hefler et al., 2016, Sifour et al., 2023).

In movement ecology, Pantala flavescens also anchors a route-optimization study of transoceanic migration. The paper models the India–Maldives–Seychelles–East Africa circuit with a Dragonfly Energetics Model and a Dijkstra-based path planner with active wind compensation. For the species parameters used, the maximum-range speed is $\mu_B < 30$42 m s$\mu_B < 30$43, the endurance is $\mu_B < 30$44 h, and the still-air range is $\mu_B < 30$45 km. Ground velocity along a chosen track satisfies

$\mu_B < 30$46

with crosswind compensation constrained by $\mu_B < 30$47. The study finds that the Somali Jet makes a direct Africa→India crossing feasible, whereas the return requires stopovers in the Maldives and Seychelles. The annual migration spans roughly $\mu_B < 30$48–$\mu_B < 30$49 km over 4–5 generations and is synchronized to precipitation windows and the shifting ITCZ (Ranjan et al., 2022).

Across these uses, DragonFly is not a single scientific object but a distributed research label attached to instruments, algorithms, models, missions, networks, and biological systems. What unifies them is not a shared method but a recurring association with distributed sensing, motion through complex environments, and fine-grained structure at scales that conventional systems often fail to resolve.