Emu: Multi-Domain Applications & Systems

• Emu is a polysemous term spanning astronomy, machine learning, infrared instrumentation, biomechanics, and macroeconomics, with each domain applying distinct methods.
• In astronomy, EMU refers to the Evolutionary Map of the Universe survey using ASKAP to achieve ~10 μJy/beam sensitivity and catalog tens of millions of galaxies.
• In machine learning and other fields, Emu encompasses multimodal generative models and NLP frameworks, while in economics it denotes the European Monetary Union.

“Emu” and “EMU” are used in contemporary research as names for several unrelated systems, programs, and concepts. The term denotes, among other things, the Evolutionary Map of the Universe radio survey in astronomy, a family of multimodal and generative models in machine learning, a multilingual sentence-embedding specialization framework, an infrared survey concept for the International Space Station, an efficient musculoskeletal simulator, and the European Monetary Union in macroeconomics (Norris et al., 2011, Sun et al., 2023, Hirota et al., 2019, Mathew et al., 2022, Modi et al., 2020, Kirrane, 2018). The shared label is therefore polysemous rather than disciplinary; interpretation depends entirely on context.

1. Nomenclature and domain-specific meanings

In astronomy, EMU stands for Evolutionary Map of the Universe, a wide-field ASKAP radio continuum survey (Norris et al., 2011). In machine learning, Emu names several Meta systems, including a multimodal autoregressive foundation model, a quality-tuned text-to-image model, and a factorized text-to-video model (Sun et al., 2023, Dai et al., 2023, Girdhar et al., 2023). In NLP, Emu is also a framework for semantic specialization of multilingual sentence embeddings (Hirota et al., 2019). Outside AI, the name appears in an ISS infrared mission concept and in EMU: Efficient Muscle Simulation in Deformation Space (Mathew et al., 2022, Modi et al., 2020). In economics, EMU conventionally denotes the European Monetary Union or European Economic and Monetary Union (Kirrane, 2018, Peeters, 2018, Kirrane, 2018).

A common source of confusion is that these usages are not variants of one research line. The astronomy EMU is a survey infrastructure and cataloging ecosystem; the machine-learning Emu systems are model families with distinct objectives and training pipelines; the macroeconomic EMU is an institutional arrangement around a single currency and common monetary authority. This suggests that “Emu” functions as a recurrent project name rather than a stable technical designation.

2. EMU as the Evolutionary Map of the Universe

The Evolutionary Map of the Universe was introduced as a wide-field radio continuum survey planned for the new Australian Square Kilometre Array Pathfinder (ASKAP). Its primary goal was to make a deep survey with rms $\sim 10~\mu\mathrm{Jy/beam}$ of the entire Southern Sky at 1.3 GHz, extending as far north as $+30^\circ$ declination, with a resolution of 10 arcsec (Norris et al., 2011). The survey was expected to detect and catalogue about 70 million galaxies, including typical star-forming galaxies up to $z\sim 1$, powerful starbursts to even greater redshifts, and AGNs to the edge of the visible Universe (Norris et al., 2011).

Later descriptions emphasize ASKAP’s role in enabling this scale. ASKAP is described as having 36 12-m antennas with phased-array feeds and an instantaneous field of view of $30~\mathrm{deg}^2$, allowing rapid wide-area imaging at roughly GHz frequencies (Vardoulaki et al., 24 Sep 2025). EMU is repeatedly characterized as a southern-sky survey reaching to $+30^\circ$ declination and targeting tens of millions of radio sources (Vardoulaki et al., 24 Sep 2025, Tang et al., 19 Jun 2025). A related cosmology-oriented characterization gives EMU-ASKAP a sky coverage of about 75% of the sky, roughly $3\pi$ steradians, with sensitivity about $10~\mu\mathrm{Jy/beam}$ and resolution about 10 arcseconds (Rahman et al., 2016).

Pilot-survey papers document the transition from design concept to survey data products. EMU Pilot Survey 1 is described as being observed at 944 MHz, covering $270~\mathrm{deg}^2$, with a native resolution of about $11\times 13$ arcsec and rms sensitivity of about $25~\mu\mathrm{Jy\,beam}^{-1}$ (Norris et al., 31 Jul 2025). A cluster-focused pilot study describes the main survey as mapping the Southern Sky at 943 MHz with 10 h per pointing, a typical synthesized beam of $+30^\circ$0, and target noise of about $+30^\circ$1 (Duchesne et al., 2024). These values reflect different stages and products rather than a single immutable observing mode.

Scientifically, EMU is framed as a census instrument for star formation, AGN activity, cosmic structure, and rare radio phenomena (Norris et al., 2011, Vardoulaki et al., 24 Sep 2025). The recurring expectation that it will “undoubtedly discover new classes of object” is not merely rhetorical in context: later EMU-related cataloging papers explicitly discuss structurally unusual populations such as Odd Radio Circles, winged systems, restarted sources, and morphologies not seen in earlier observations (Norris et al., 2011, Tang et al., 19 Jun 2025, Norris et al., 31 Jul 2025).

3. Source finding, cataloging, and radio morphology

Because EMU-scale source counts preclude manual catalog construction, automated source finding became a central methodological problem early in the project. The ASKAP/EMU Source Finding Data Challenge tested eleven source-finding tools submitted by nine teams on simulated ASKAP-like images (Hopkins et al., 2015). Its headline result was that completeness is close to 100% at about $+30^\circ$2 and drops to around 10% by about $+30^\circ$3, while reliability is typically close to 100% at about $+30^\circ$4 but varies strongly between finders at lower signal-to-noise (Hopkins et al., 2015). The paper emphasizes the standard completeness–reliability trade-off and uses the outcome to guide improvements to Selavy, the ASKAPsoft prototype source finder (Hopkins et al., 2015).

Extended and morphologically complex sources proved to be the main residual difficulty. A later RGZ EMU framework paper states that EMUCAT handles only about $+30^\circ$5 of EMU sources reliably and that EMU is expected to contain about $+30^\circ$6 million extended radio sources (Tang et al., 19 Jun 2025). To address this, Radio Galaxy Zoo: EMU (RGZ EMU) combines citizen science, machine learning, expert validation, and multi-wavelength cross-matching (Tang et al., 19 Jun 2025). In its Phase I design, the workflow started from more than 200,000 entries in the EMU PS 1 Selavy component catalog, kept only cutouts with complexity $+30^\circ$7, reducing the pool to 37,578 objects, and then applied a major axis $+30^\circ$8 arcsec criterion to define a 6,230-image citizen-science sample (Tang et al., 19 Jun 2025). The interface asked volunteers to perform Radio Source Assembly, Radio Morphology Tagging, and optional Talkboard commentary using ASKAP, WISE, and DES images at multiple angular scales (Tang et al., 19 Jun 2025). An early report gave 1,435 citizen scientists and $+30^\circ$9 classifications since launch on 8 July 2024, while a later Phase I paper reported more than 2,500 volunteers and 97,000 classifications (Tang et al., 19 Jun 2025, Vardoulaki et al., 24 Sep 2025).

The morphological complexity of EMU sources is also evident in specialist catalogs. “EMU and the DRAGNs I: A Catalogue of DRAGNs” presents 3557 double radio sources associated with active galactic nuclei from EMU-PS1, each extracted and identified by eye, tagged morphologically, and measured for size and flux (Norris et al., 31 Jul 2025). The catalog uses the Fanaroff–Riley ratio $z\sim 1$0, where $z\sim 1$1 is the separation between the brightness maxima on opposite sides of the source and $z\sim 1$2 is the total extent, with

$z\sim 1$3

and an uncertainty-aware FRX class defined using $z\sim 1$4 and $z\sim 1$5 (Norris et al., 31 Jul 2025). Reported populations include 1410 FR2, 238 FR1, 42 HyMoRS, 696 linear triple sources, 34 one-sided sources, 243 bent-tail sources, 17 head-tail candidates, 20 double-double sources, and 15 WTF sources (Norris et al., 31 Jul 2025). The tag-based philosophy explicitly allows one source to carry more than one morphology label (Norris et al., 31 Jul 2025).

No single automatic method recovers all extended sources. A comparative EMU-G09 study applied DRAGNHunter, coarse-grained complexity, and RG-CAT and found that only 375 sources were identified by all three methods (Barnes et al., 11 Mar 2026). DRAGNHunter favors classical double-lobed systems, coarse-grained complexity highlights morphologically rich emission, and RG-CAT tends toward larger and brighter radio galaxies (Barnes et al., 11 Mar 2026). This suggests that EMU extended-source cataloging is intrinsically ensemble-based rather than reducible to one universal finder.

4. Cosmology, redshift inference, and diffuse cluster emission

EMU’s survey geometry and source density make it a cosmology instrument as well as a source catalog. A technical cosmology analysis treats EMU as a flagship radio continuum survey for galaxy over/under-density maps, emphasizing that confusion, position accuracy, shot noise, masking, and sky coverage control the detectability of weak signals such as the late-time ISW effect (Rahman et al., 2016). In that treatment, shot noise is written simply as

$z\sim 1$6

where $z\sim 1$7 is the number of sources per steradian (Rahman et al., 2016). The same paper argues that EMU + WODAN is substantially more powerful for ISW work than EMU Early Science alone because large sky coverage is as important as depth on the largest angular scales (Rahman et al., 2016).

A more recent EMU cosmology result uses clustering redshift inference. “EMU: Cross-correlating EMU Pilot Survey 1 with Dark Energy Survey to constrain the radio galaxy redshift distribution” uses EMU-PS1 radio data over about $z\sim 1$8 at 943 MHz, with angular resolution 11–18 arcsec and rms depth 25–30 $z\sim 1$9, and applies a flux cut of $30~\mathrm{deg}^2$0 to obtain a final sample of

$30~\mathrm{deg}^2$1

radio sources (Saraf et al., 9 May 2025). Cross-correlation with DES MagLim tomography is used to infer the radio $30~\mathrm{deg}^2$2, and the recovered distribution is reported to fit the data much better than current simulated SKADS and TRECS models, peaking at a much higher redshift, around

$30~\mathrm{deg}^2$3

with best-fit EMU bias

$30~\mathrm{deg}^2$4

(Saraf et al., 9 May 2025) This suggests that simulation-based radio redshift priors can be substantially mis-centered for EMU-like samples.

The same cosmological logic appears in radio–optical cross-survey work with Euclid. An EMU–Euclid study reports the first measurement of the clustering cross-spectrum between radio-continuum sources in the EMU Main Survey and galaxies from the Euclid Q1 release, using two radio-source catalogs built with different source finders (Piccirilli et al., 27 Nov 2025). The abstract reports detection of the cross-correlation signal at above $30~\mathrm{deg}^2$5 and states that the two measured cross-spectra are in excellent agreement, implying robustness against the choice of source-finding algorithm (Piccirilli et al., 27 Nov 2025).

EMU is also well suited to low-surface-brightness cluster science. A pilot search for diffuse, non-thermal radio emission in 71 PSZ2 clusters from archival ASKAP observations detected 21 radio halos—12 for the first time, excluding an additional six candidates—11 relics in seven clusters, with six first-time detections, and 12 other unclassified diffuse radio sources (Duchesne et al., 2024). Extrapolating to the 858 PSZ2 clusters expected to be covered by the full survey, the paper predicts up to $30~\mathrm{deg}^2$6 radio halos and $30~\mathrm{deg}^2$7 radio relics (Duchesne et al., 2024). The diffuse-emission search relied on both $30~\mathrm{deg}^2$8-plane filtering and image-plane angular-scale filtering, reflecting EMU’s dual role as both survey and reprocessing substrate (Duchesne et al., 2024).

5. Emu in multimodal and generative machine learning

In machine learning, the name Emu refers to several distinct systems rather than one model. The broadest is “Emu: Generative Pretraining in Multimodality”, a 14B-parameter foundation model initialized from EVA-CLIP, LLaMA-13B, and Stable Diffusion v1.5 (Sun et al., 2023). Its central idea is a unified autoregressive objective over interleaved text and visual embeddings: $30~\mathrm{deg}^2$9 where the next element may be a text token or a visual embedding (Sun et al., 2023). The model uses image-text pairs from LAION-2B and LAION-COCO, interleaved image-text documents from MMC4, video-text pairs from WebVid-10M, and interleaved video-text data from YT-Storyboard-1B, which is built from 18 million YouTube videos and about 1.8 billion storyboard images (Sun et al., 2023). Emu is reported to support image captioning, VQA, video QA, text-to-image generation, in-context multimodal generation, and instruction-following via Emu-I (Sun et al., 2023).

A separate paper, “Emu: Enhancing Image Generation Models Using Photogenic Needles in a Haystack”, names a text-to-image model whose main contribution is quality-tuning (Dai et al., 2023). The system pre-trains a latent diffusion model on $+30^\circ$0 billion image-text pairs and fine-tunes it with exactly 2000 exceptionally high-quality images with manually written captions (Dai et al., 2023). The reported outcome is a strong human-preference gain: 82.9% win rate over its own pre-trained-only counterpart on visual appeal, and 68.4% and 71.3% preference over SDXLv1.0 on PartiPrompts and Open User Input, respectively (Dai et al., 2023). The paper argues that a small curated set can restrict generation to a high-quality subset of image space without sacrificing concept coverage (Dai et al., 2023).

The video variant, “Emu Video: Factorizing Text-to-Video Generation by Explicit Image Conditioning”, factorizes generation into two stages,

$+30^\circ$1

where a text-conditioned image is generated first and a video is then generated from the text and the image (Girdhar et al., 2023). The model uses a pretrained Emu text-to-image system, a U-Net with about 2.7B frozen spatial parameters and 1.7B trainable temporal parameters, and is trained on 34M licensed video-text pairs (Girdhar et al., 2023). The paper identifies a zero terminal-SNR noise schedule and multi-stage multi-resolution training as critical design choices, and reports human-preference rates of 81% over Imagen Video, 90% over PYOCO, 96% over Make-A-Video, and 96% preference over VideoComposer for image animation (Girdhar et al., 2023).

These three systems share a name but not an objective. One is a multimodal autoregressive generalist; one is a quality-aligned latent diffusion image generator; one is a factorized text-to-video model. The overlap is conceptual rather than architectural.

6. Other technical systems called Emu

In multilingual NLP, Emu is a lightweight framework for semantic specialization of multilingual sentence embeddings (Hirota et al., 2019). It uses a multilingual encoder $+30^\circ$2, semantic classifier $+30^\circ$3, and language discriminator $+30^\circ$4, with a classifier loss

$+30^\circ$5

and adversarial training objective

$+30^\circ$6

(Hirota et al., 2019) The purpose is to correct the textual-similarity bias of off-the-shelf multilingual embeddings while preserving multilingual transfer. The paper reports that the specialized embeddings outperform the prior state of the art on cross-lingual intent classification using only monolingual labeled data (Hirota et al., 2019).

In infrared instrumentation, Emu is a proposed ISS-hosted near-infrared sky-survey telescope designed for TDI-like / drift-scan imaging without active pointing control (Mathew et al., 2022). It is a 6-month mission concept centered on the $+30^\circ$7 water-absorption band, using the Leonardo SAPHIRA eAPD array and ANU Rosella electronics (Mathew et al., 2022). The mission is described as a 6U CubeSat-form-factor payload with a compact Cassegrain telescope, 10 arcsec pixels, a 25 Hz science frame rate, and sky access between $+30^\circ$8 and $+30^\circ$9 declination, covering about 78% of the sky (Mathew et al., 2022). Its science driver is oxygen-abundance inference in cool stars through the strength of the $3\pi$0 H$3\pi$1O band (Mathew et al., 2022).

In computational mechanics and graphics, EMU abbreviates Efficient Muscle Simulation in Deformation Space, a quasi-static finite-element-style framework for bulk musculoskeletal systems (Modi et al., 2020). Its core innovation is to use per-tetrahedron deformation gradients $3\pi$2 as primary unknowns and to enforce geometric consistency through an As-Continuous-As-Possible (ACAP) energy

$3\pi$3

(Modi et al., 2020) The method is designed to support heterogeneously stiff meshes and arbitrary constitutive models, allowing soft muscles, stiff tendons, and stiffer bones in one unified system (Modi et al., 2020). Reported engineering outcomes include handling a 600k-tetrahedron muscle model on a 16GB laptop and achieving up to $3\pi$4 speedups over state-of-the-art FEM on medium and large meshes (Modi et al., 2020).

These systems have no substantive relation to one another beyond nomenclature. Their coexistence under the same label illustrates how research names often propagate independently across communities.

7. EMU in macroeconomics and European integration

In economics, EMU denotes the European Monetary Union or European Economic and Monetary Union, the Maastricht-era project of irrevocably fixed exchange rates, a single currency, and a common monetary authority (Kirrane, 2018, Kirrane, 2018). The literature summarized here treats EMU as a trade-off between the benefits of exchange-rate stability, lower transaction costs, and greater policy credibility, and the costs of lost national monetary autonomy and more difficult adjustment to asymmetric shocks (Kirrane, 2018, Peeters, 2018).

One major theme is institutional design. Kirrane’s analysis stresses three choices under Maastricht: a single central bank, central bank independence with price stability as the primary objective, and national fiscal autonomy subject to discipline and surveillance (Kirrane, 2018). The same paper highlights the Treaty’s reference values of deficits below 3% of GDP and public debt below 60% of GDP, while also noting criticism that these thresholds can be rigid and not always economically sensible in recession (Kirrane, 2018). A related analysis of stability in EMU argues that the Maastricht debt and deficit ceilings were motivated by monetary credibility rather than short-run stabilization and presents the later Stability Pact as a compromise regime with sanctions for excessive deficits, in the form of unpaid deposits that can become fines, combined with exemptions for exceptional events or severe recession (Peeters, 2018).

Historical treatments place European EMU in a longer lineage of monetary unions. One paper compares nineteenth-century German, Italian, and Japanese monetary integration with the Latin Monetary Union, Scandinavian Monetary Union, West African Monetary Union, and the European Monetary System, arguing that states enter such arrangements when expected national benefits exceed expected national costs (Kirrane, 2018). Political integration, leadership by a dominant state, small group size, and credible monetary backing are recurrent success conditions in that account (Kirrane, 2018). This suggests that the acronym EMU in economics names not a technical model but a historically contingent institutional bargain around sovereignty, credibility, and adjustment.

Across these economic papers, a recurrent controversy concerns whether EMU should prioritize anti-inflationary credibility or macroeconomic stabilization. The literature does not treat Maastricht discipline as self-evidently optimal; it instead frames European EMU as a durable but contested balance between fiscal rules, central bank independence, and the need to absorb asymmetric shocks (Peeters, 2018, Kirrane, 2018).