A New View of Observed Galaxies through 3D Modelling and Visualisation
Tim Dykes, Claudio Gheller, Bärbel S. Koribalski, Klaus Dolag, Mel Krokos
AA New View of Observed Galaxies through 3D Modelling and Visualisation
Tim Dykes , Claudio Gheller b , B¨arbel S. Koribalski c , Klaus Dolag d , Mel Krokos e a HPE HPC/AI EMEA Research Lab, Broad Quay House, Broad Quay, Bristol, U.K. b Institute of Radioastronomy, INAF. Via P. Gobetti, 101 40129 Bologna, Italy c CSIRO Astronomy and Space Science, Australia Telescope National Facility, P.O. Box 76, Epping, NSW 2121, Australia. d Universit¨ats-Sternwarte M¨unchen, Scheinerstr.1, D-81679 M¨unchen, Germany. e School of Creative Technologies, University of Portsmouth, Portsmouth PO1 2DJ, U.K.
Abstract
Observational astronomers survey the sky in great detail to gain a better understanding of many types of astronomicalphenomena. In particular, the formation and evolution of galaxies, including our own, is a wide field of research. Threedimensional (spatial 3D) scientific visualisation is typically limited to simulated galaxies, due to the inherently twodimensional spatial resolution of Earth-based observations. However, with appropriate means of reconstruction, suchvisualisation can also be used to bring out the inherent 3D structure that exists in 2D observations of known galaxies,providing new views of these galaxies and visually illustrating the spatial relationships within galaxy groups that arenot obvious in 2D. We present a novel approach to reconstruct and visualise 3D representations of nearby galaxies basedon observational data using the scientific visualisation software Splotch. We apply our approach to a case study of thenearby barred spiral galaxy known as M83, presenting a new perspective of the M83 local group and highlighting thesimilarities between our reconstructed views of M83 and other known galaxies of similar inclinations.
Keywords: galaxy modelling, visualisation
1. Introduction
The study of galaxy formation and evolution is a wideresearch field in astronomy, ranging from multi-frequencyobservations of the Milky Way and nearby galaxies to dis-tant galaxy groups and clusters [33, 18, 26], complementedby cosmological galaxy simulations from the Big Bang tothe present day [32, 44]. While specific galaxies can onlybe observed from a single viewpoint on (or near) Earth,their intrinsic 3D shapes are well known from all-sky sur-veys, such as the Sloan Digital Sky Survey (SDSS) [58],containing millions of objects at all possible orientationangles. As an example, Figure 1 highlights the contrast-ing views of two similar spiral disk galaxies seen at verydifferent angles (represented by their inclination angle i ):the barred galaxy M 83 ( left ) is seen nearly face-on ( i ∼ right ) is seen nearlyedge-on ( i ∼
90 degrees).Detailed observations of galaxies over a large range offrequencies (e.g. optical, radio, infrared and ultraviolet)allow the measurement and derivation of many differentproperties, together enabling us to study galaxy morphol-ogy, kinematics, composition, mass, age, and formationhistory. Conversely, simulation of astronomical objects hasbeen an active research topic in astronomy for many years[40, 54] with an aim to gain a better understanding ofphysical processes occurring over timescales so long, and Corresponding author: [email protected] across spatial scales so massive, that we cannot hope toobserve them in real-time. In both cases, visualisationhas been used extensively to help understand, analyse,and disseminate the results of such efforts, as discussedin Section 2. In the context of the study of galaxies, vi-sualisations are typically created from multi-dimensionaldata fields generated by numerical simulations and used tostudy properties of the source data; the focus is on present-ing data in a manner facilitating intuitive comprehensionrather than showing a visually realistic galaxy representa-tion. Conversely, realistic 3D spatial visualisation of astro-nomical objects based on observations requires physicallyrobust reconstruction methods paired with visualisationtools that support high quality rendering.This paper presents a novel methodology to reconstructand visualise particle-based 3D models of nearby galaxiesbased on (a) observed multi-wavelength images and (b)detailed kinematic information. Our objectives are to: (1)create realistic views of galaxies including from viewpointsnot otherwise possible to observe, (2) explore the validityof derived spatial 3D models of such galaxies, and (3) al-low enhanced visual analysis of the 3D galaxy morphology(stars, gas and dust). The methodology is implementedwithin the open-source 3D scientific visualisation packageSplotch [11], enabling the reconstruction, visualisation,and exploration of individual, or groups of, recognisablegalaxies. We expect that our pipeline may be utilised for Preprint submitted to Astronomy and Computing February 8, 2021 a r X i v : . [ a s t r o - ph . I M ] F e b cientific communication and outreach by generating im-mersive and realistic movies in 3D of known galaxies, al-lowing the non-expert viewer to grasp the connection be-tween 2D observations and the real 3D structure of suchobjects. We also separate out components within that 3Dstructure, highlighting for example the large extent andwarped morphology of the gaseous component that is of-ten forgotten by both experts and non-experts alike. Wehope that through physically realistic reconstruction andrendering this pipeline may also be used to complimenttypical statistical analyses via objectives (2) and (3), al-lowing the astronomer to build better mental images ofboth the 3D structure of their object of study, and bet-ter comprehend spatial relationships between companionobjects.As a case study for our work, we reconstruct and visu-alise the nearby barred spiral galaxy known as M 83, for itsdesignation in the Messier catalogue, or as NGC 5236 inthe New General Catalogue (NGC) of Nebulae and Clus-ters of Stars. As one of the closest and brightest spiralgalaxies in the sky, M 83 has a large collection of highresolution and multi-wavelength data publicly available,including H i i Sur-vey (LVHIS; Koribalski et al. 2018). Furthermore, M 83 isparticularly well suited for visualisation due to its nearlyface-on inclination and a large warped H i disk, making itpossible to use our sophisticated 3D reconstruction to vi-sualise its likely edge-on appearance. The numerous dwarfcompanion galaxies surrounding M 83 have also been mod-elled, allowing us to create a 3D visualisation of the galaxygroup.The structure of this paper is as follows: Section 2provides an overview of related work. Section 3 presents abrief introduction to the techniques for galaxy observationand kinematic modelling that underpin our galaxy recon-struction approach. Section 4 introduces a new techniquefor reconstructing the physical structure of observed galax-ies. Section 5 describes our galaxy visualisation process,with results presented in Section 6. Section 7 aims to eval-uate the extent to which the described approach achievesthe goal of realism. Section 8 presents concluding remarksand suggests future directions.
2. Related Work
Visualisation is an integral part of astronomical anal-ysis, helping domain scientists to explore their data, iden-tify problems or areas of interest for further analysis, aswell as to convey important concepts and results duringdissemination of their work [20, 51, 13]. We first providea brief contextual overview of visualisation approaches ofsimulated and mock data, and then focus on relevant re-search to the primary topic of the paper: multi-frequencyand observed data reconstruction and visualisation. http://bf-astro.com/index.htm Figure 1: Optical images of two very similar spiral disk galaxies seenat different inclination angles. The face-on view of M 83 ( left ) re-veals the full complexity of its stellar spiral structure mixed withdust and young star-forming regions. In contrast, the edge-on viewof NGC 4565 ( right ) reveals an old central bulge and strongly obscur-ing dust lanes seen against the thin stellar disk. — Images courtesyof
NASA, ESA, and the Hubble Heritage Team (STScI/AURA) and
Robert Franke , respectively. — North is up and East to the left. Cosmological simulations allow astronomers to repli-cate physical processes to the best of their collective knowl-edge, simulating billions of years of evolution on timescalesmany orders of magnitude faster than real-time. Such sim-ulations allow us to observe the behaviour of matter on thescale of the full evolution of the known universe in just afew days of computational processing using state of the artsimulation codes and supercomputers [44, 49]. The higherthe resolution of a simulation, the easier it is to studydiscrete astronomical objects, such as single galaxies, asthey evolve from simple clouds of matter to complex dy-namic systems that contain hundreds of billions of stars[19, 37, 43]. 3D visualisation in this context (e.g. [10])typically relies on the already existing 3D spatial struc-ture from the simulation data. Our work, by contrast, isbased on reconstruction and visualisation of 3D structurebased on 2D observed images. We place emphasis on phys-ical realism in our visualisation, by exploiting scientificallyrobust parameters where possible in our modelling and vi-sualisation (discussed further in Section 7), other workshave focused on achieving visual realism of scientific datathrough advanced rendering techniques, exploiting render-ing software typically used in movie production to generateimpressive cinematic visualisations of astrophysical data[36, 6].A related research area in Astronomy involves the cre-ation of so-called mock galaxy catalogues . Such catalogues,built from simulated data, consist of a structured col-lection of galaxies used to support development of scien-tific pipelines for the collection, validation, and analysis,of large scale observational surveys [8]. This process in-cludes mock imaging , or the generation of synthetic ob-servational images of simulated data. Such images canbe used for the setup and tuning of observational instru-ments, and allow image-based comparison between resultsof theoretical simulation and physical observation for val-2dation purposes (e.g. [7, 59]). Tools such as SkyMaker[2] and Phox [4] can be considered realistic visualisationtools that aim to generate 2 dimensional representations ofsimulated galaxies in specific frequency ranges, with manybespoke features such as replication of optical effects fromtelescopic instrumentation. In general our work targets3D visualisation and exploration, as opposed to allowingquantitative comparison with observed images, and so wedo not exploit this type of tool.There are a variety of works studying 3D visualisationof multi-dimensional astronomical data consisting of oneor more spatial dimensions plus frequency. One approachis to consider the frequency domain as the third spatial di-mension to use typical 3D visualisation methods, as in [28].Whilst effective for studying frequency domain data, thisis not sufficient for a realistic representation in 3 spatialdimensions. In some cases, a velocity field can be derivedfrom frequency data for a closer approximation to phys-ical structure, as in [52, 55, 45]. Other approaches aimto reconstruct the proper 3D spatial structure for visual-isation. Astronomical tomographic techniques for recon-structing 3D structure have been used to good effect, forexample in [27], who reconstruct and visualise areas of theLyman- α forest, large clouds of gas that absorb Lyman- α radiation emitted from quasars.The work of [35] exploits a 3D modelling techniquefor the Orion nebula based on infrared and optical ob-servational data [56] combined with volume rendering tovisualise stars and emission nebulae. The work of [29,30] presents a comprehensive modelling and visualisationpipeline for planetary nebulae, based on a novel ConstrainedInverse Volume Rendering technique to reconstruct gasdistributions from a series of optical images; this techniquefurther underpins the work of [22], who reconstruct andvisualise spiral galaxy M 81 based on optical and infraredimages. The software tool Shape[50] builds upon this mod-elling approach to provide an interactive construction andanalysis tool for planetary nebulae, used in several relatedworks focusing on reconstruction of planetary nebulae (e.g.[17, 1]). The SlicerAstro [45] extension to scientific visuali-sation software 3DSlicer [14] allows users to visualise (H i )data in 3D, two spatial dimensions and a third velocitydimension. This is combined with interactive features forexploring such data and visualisation of tilted-ring basedkinematic models (as discussed further in Section 3).The scope of our work is reconstruction and visuali-sation of observed galaxies based on multi-frequency datafrom observation catalogues. We present a methodologyfor galaxy reconstruction and visualisation that exploits anovel particle-based approach to reconstruct the 3D struc-ture of spiral galaxies from multi-frequency observed im-ages and kinematic models, utilising the high performancevisualisation software Splotch for high quality visualisa-tion. Previous works have covered a wide range of fre-quency bands, such as [28], but are focused on the visualexploration of data in those frequency bands rather thanspatial reconstructions based on such data. The work of [22] has similarly modelled spiral galaxies, however usedfewer data sources and focused only on the stellar disk,achieving a less comprehensive galaxy view than our work.[50] has developed a full interactive construction and anal-ysis tool, however the focus is on other types of astronom-ical object such as astronomical nebulae, and the tech-niques do not directly apply to spiral galaxies.As will be described in following sections, our approachis able to: exploit a wider range of source data than relatedwork, allowing us to incorporate an absorptive dust laneand the extended distribution of atomic hydrogen typicallyseen around spiral galaxies; utilise kinematic models to re-cover an approximation of 3D shape for spiral galaxies;and combine multiple objects into a single scene repre-senting a group or cluster of galaxies. We also modify theSplotch software to better support such visualisations byextending the transfer function to treat emission and ab-sorption fully independently to support dust occlusion inour visualisation.
3. Observational structure and dynamics of galax-ies
This section presents a brief overview of the astronom-ical approach to inferring structure and kinematics of ob-served galaxies, which we will exploit for 3D reconstruc-tion. We introduce two key domain-specific concepts, re-lating to the varied structure of galaxies observable acrossthe electromagnetic spectrum (Section 3.1), and the con-cept of tilted-ring fitting for galactic hydrogen disks inspiral galaxies (Section 3.2).
In terms of physical structure, a typical spiral galaxy(like M 83) consists of a bright stellar disk, an extended andwarped gaseous hydrogen disk, a central elliptical bulge,and a dark-matter dominated halo. The stellar compo-nent is dominated by prominent spiral arms, which alsocontain large amounts of dust (see Fig. 1). A comprehen-sive understanding of this structure requires observationsacross multiple wavelengths. For example, the bright stel-lar body is typically captured at optical, infrared (IR) andultraviolet (UV) wavelengths whilst cold atomic hydrogengas (H i ), which is observed in spiral disk galaxies whereit typically extends far beyond the bright stellar disk, isonly observable with radio telescopes. Images of the stellarbody observed at optical, IR, and UV wavelengths can in-form us of different types of galactic population. Emissionfrom the youngest, most massive, stars is mostly seen inUV, whilst older stars dominate optical observations, anddifferent ranges of these spectra further segregate stellarpopulations by structure, age, and other physical proper-ties [33, 18]. IR emission in the 8 µ m range is typically usedto identify and study interstellar dust [21], which absorbsradiation other parts of the spectrum and re-emits it inIR.3o visualise galaxies of different sizes and types, an un-derstanding of known relations can be beneficial. For ex-ample, the total dynamical mass (a quantity derived fromvelocity observations) M dyn of a disk galaxy scales with itsrotational velocity and its radius R , i.e. M dyn ∝ v × R .Dwarf galaxies are low-mass, slow rotators, typically witha thick disk tending towards spheroidal shapes, while themost massive spiral galaxies have very large, thin disks.We distinguish between elliptical and disk galaxy compo-nents in Section 4.3, where different generating functionscan replicate disk- or elliptical- shape galaxies.For visualisation purposes, other galaxy properties suchas the disk length and height of stellar and gas compo-nents can be determined by analysing a set of observations.It is often difficult to directly measure the height or diskthickness from 2D observed surface density maps, due toprojection effects. However, in a study of nearby edge-ongalaxies, [38] found that the rotational velocity v rot relatesto the disk thickness z height ; the faster galaxies rotate, thethinner their disks ( z height ∝ /v rot ). As such, modellingthe galaxy kinematics is essential in determining their 3Dshape, as discussed in Section 3.2. The hydrogen (H i ) disks of spiral galaxies are gener-ally found to be (a) much more extended (by a factor 2–3)than the bright stellar disk and (b) warped as illustratedin Figure 5. Such warps, which can range from a few de-grees to tens of degrees, typically start at the edge of thebright stellar disk and extend outwards. To incorporatethis warped shape in our reconstruction, we must approx-imate it in 3D.The de-facto approach for obtaining a best-fit shapefor a H i gaseous disk is via a tilted-ring analysis , a long-standing means of investigating kinematic galaxy structure[48] based on modelling the H i velocity field of galaxiesfor which we have high resolution observations (see, e.g.,[39]). A tilted-ring analysis generates a tilted-ring model,represented by the galaxy inclination ( i ), position angle( P A ), thickness ( Z ), and rotational velocity ( v rot ), as afunction of radius ( r ), and allows us to derive the 3D shapeof H i -rich spiral galaxies.In this work, we utilise a tilted-ring model generatedwith the TiRiFiC software [24]. Our modelling approachrequires one ascii text file per component consisting of theinclination angle, the position angle and the disk heightas a function of radius; for future use we also include therotational velocity as a function of radius. Such ascii filescan be created from the output of any tilted-ring mod-elling software. For the purpose of 3D visualisation, tilted-ring models of regularly rotating galaxies are most useful.As the gaseous disks of galaxies are typically warped andmuch larger than the bright stellar body (factor 2–3), werequire high resolution H i
4. Particle Based Modelling of Observed Galaxies
In this section we describe our approach for construct-ing a particle-based 3D galaxy representation based on ob-servational data and a tilted-ring model as introduced inSection 3. Our representation consists of multiple galacticcomponents: • Stellar population • Diffuse hydrogen gas • Dust • Galactic bulge • Globular clusters surrounding the galaxy • Diffuse stellar, or dark matter, haloWe first collect and pre-process a set of observationalsource data and a tilted-ring model that will inform thedensity and spatial distribution of particles. Then, for eachgalaxy component, we construct a 3D distribution of tracerparticles. In this context, a particle is considered a pointsource in 3D space with a series of inherent properties,such as a radius, a colour, and an intensity of light emis-sion or absorption. The construction is implemented foreach of our galactic components separately and is, wherepossible, based on the available observed and/or derivedstructural data. Components may also be tuned empiri-cally using additional parameters which can be obtainedfrom well known observed relations for the different galaxytypes or components to present a more accurate represen-tation of, for example, the thickness of the gas layer inthe stellar disk (as in Section 4.3). Finally each of thesetracer particles are coloured from observational images toreproduce the visual appearance as seen in the observa-tions, and the galaxy components are combined to formthe complete model.The reconstruction and visualisation methodology isstructured as a pipeline beginning with observed imagesof known galaxies, and ending with 3D visualisations ofreconstructed, particle-based, galaxy models; the pipelineis illustrated by the block diagram shown in Figure 2. Thispipeline is used for both spiral and elliptical galaxies, withparameter files used to distinguish which inputs are re-quired. The following sub-sections provide an overview ofeach stage of the modelling pipeline with associated in-puts and outputs, illustrated via our case study galaxyM 83. Section 5 will then detail the visualisation process.4
ITS to BIN Galaxy Modeller Display
3D Galaxy Visualization Pipeline
Input DataPipeline Components SplotchCombine Galaxies
Cleaned FITS imagesColour choices
Optional path.sceneGadget fileNgcxxxx_0000
Input Output
Optional Step to place multiple galaxies in one scene
Full Pipeline : Images to Movie path.scenecreate_galaxy.par
Output Data
Img00.tgaImg01.tgaImg02.tga…Media display (.tga)Movie display (.avi)
Image Preparation
Cleaned FITS images Binary image masks Binary image masks Gadget fileNgcxxxx_0000
Pipeline Direction
Figure 2: A block diagram representation of the galaxy modelling pipeline starting from observed images of a specific galaxy, and resultingin visualised images of the reconstructed 3D galaxy model. Each of the stages, inputs, and outputs are described further in the text.
The first step of the pipeline is collection and prepara-tion of the source data, which later will be used to informthe physical structure and colouring of different galaxycomponents; observed images are collected in the optical,UV, IR, and radio wavelengths, alongside 3D structuraldata from a tilted-ring analysis where possible. A varietyof data is obtained to capture the diverse galactic popula-tions as discussed in Section 3, exploiting existing surveydata based on the availability for a specific galaxy.In the case of M 83, optical H α - and R -band imagesare used from the Survey for Ionization in Neutral GasGalaxies (SINGG) [33] and near- and far- U V images fromthe Galaxy Evolution Explorer (GALEX) [18] to informthe stellar component. To represent the structure of thegas component, radio interferometric H i intensity mapsat up to three different resolutions are created from theAustralia Telescope Compact Array (ATCA) data. Mostimportantly, the H i data is used to determine the extent,shape, and kinematics of galaxies, described in the nextsection. The “Local Volume H i Survey” [26] providesATCA H i data for nearly 100 nearby galaxies ( D < . The dust dis-tribution, which is to varying degrees already part of the observed stellar component, is informed by 8 µ m images (asin [21]) from the Spitzer Infrared Array Camera (IRAC)instrument [9], which also provides improved structure tothe spiral arms. Images are used from freely available fromexisting surveys, and so initial collection is quickly accom-plished.The multi-wavelength images are then pre-processedand organised as a set of FITS [41] files. The FITS (Flex-ible Image Transport System) format is an open standardcommonly used in astronomy for storage, transmission andprocessing of data, typically in the form of 2D or 3D im-ages, or tables. Pre-processing is a more involved task thaninitial collection, requiring an experienced astronomer toclean, normalise, orient, and scale images. For example,optical images of galaxies contain foreground stars andbackground galaxies that need to be removed (or cleaned )using domain-specific software (e.g. [3]). This stage isimportant as extraneous sources will introduce incongru-ous artefacts in the final 3D model. The normalisation,orientation, and scaling of images is required to ensureeach galaxy component is modelled accurately relative toeach other component - this task is carried out with thesame rigour as for typical image analyses. Image intensityranges are then clipped to reduce the dynamic range of pix-els. Source images are typically FITS formatted with pix-els of high dynamic range, these cannot be accurately rep-5 igure 3: The cleaned FITS images of M 83 used for reconstruction,in ultraviolet, optical, infrared, and radio bands. Top left:
GALEXnear UV-band (UV stellar distribution).
Middle left:
GALEX farUV-band (UV stellar distribution).
Bottom left:
SPITZER IRAC8 µ m-band (Dust distribution). Top right:
SINGG H α -band (Opticalstellar distribution). Middle right:
SINGG R -band (Optical stellardistribution). Bottom right:
ATCA radio-band at medium resolution(Combined with other resolutions for diffuse (H i ) distribution). resented in the colour depth (or radiometric resolution) ofimages typically used in 3D visualisation. An experiencedastronomer must choose suitable maximum and minimumcutoff values to represent the respective galaxy component,a maximum that is too high will result in over-saturation,an overpopulated model, whilst a minimum that is toolow will result in under-saturation, a sparsely populatedmodel. Figures 3 and 4 show the collection of images usedfor M 83 after pre-processing.The cleaned FITS images are then input to the FITSto BIN pipeline stage, consisting of a preprocessing scriptthat extracts, clips, and colours the FITS image pixels.The script is given, as input, colour values as RGB tripletsper input image, these colour values are combined witha per-pixel intensity and scaled to 8 bit RGB pixels foroutput as a binary image mask for the galaxy modeller.Clipping values may also be provided here to further clipminimum and maximum intensities to tune the saturationof the model. Colour values are chosen typically to high-light the different components, and if possible to matchexisting composite observed images.A tilted-ring model is used to inform the 3D structureof the galaxy disk. For M 83, we use a model generatedwith the TiRiFiC software, which provides a set of con-centric ellipses of varying radii r and thickness Z , with i , r v rot Z i P A Table 1: A minimal list of tilted-ring model values for M 83, visuallyrepresented in Figure 5. Each concentric ring is defined by a radius r ,rotational velocity v rot , thickness Z , inclination angle i , and positionangle P A . For brevity only 15 representative rings are shown, thefull model contains ∼
300 rings and was generated by Peter Kamphuisusing the TiRiFiC tilted-ring fitting software [24]. ~50 KPC
Figure 4: M 83: cleaned, preprocessed, and coloured FITS imagesof the warped outer disk in high, medium, and low spatial resolu-tions. The differing resolutions are defined by a weighting appliedwhen extracting images from the raw data, allowing to tune the ex-tent of detected emission.
Bottom left:
Low, capturing the diffusedemission.
Bottom Middle:
Medium, capturing most of the details.
Bottom Right:
High, capturing the sharp features.
Top middle:
Combined resolution, approximating the high dynamic range of theoriginal data. igure 5: A visual illustration of the TiRiFiC tilted-ring model forM 83, based on the data in Table 1. P A , and v rot values as shown in Table 1. This model, vi-sualised in Figure 5, effectively describes the warped diskstructure of M 83, and is also provided as input to thegalaxy modeller. This section describes the construction of the spatialdistribution of particles for each galactic component viathe galaxy modeller pipeline stage. The modelling is im-plemented in C++, using key-value ASCII text files forparameter inputs. Throughout this section, coordinates( x, y, z ) are used relative to the galactic plane, such that x, y are in-plane and z is axial (normal to the galacticplane).The stellar, diffuse gas, and dust particle populationsof the disk are based on the tilted-ring kinematic modeldescribed in Section 3.2. Firstly, a 3 dimensional distri-bution of N p seed particles is generated in an unweightedrandom way on the galactic plane at z = 0, defined by theconcentric tilted-ring model. This distribution of particlesin object space is projected onto the image plane of the cor-responding observational image, Obs λ . Each seed particle P is assigned an intensity I s as a function of the intensity I p of the image pixel with which it intersects, along witha scalar or three-component RGB s colour value definedby the binary image masks generated in Section 4.2. Anintensity threshold is defined as I t , defaulting to 0, andall seed particles with I s < I t , are culled, i.e. those thatlay in an area of the disk with no emission seen in image Obs λ . This creates a flat disk in the warped galactic planewith a distribution of seed particles matching the observedimage.The disk thickness is constructed by generating pointcloud distributions around the seed particles. For each ofthe remaining particles, a point cloud of size N star is gener-ated according to a Gaussian distribution defined by σ star ,and N star . Each spawned particle inherits the intensity I s and colour values RGB s of the seed particle, and is fur-ther characterised by a smoothing length h , a term fromastrophysical simulation equivalent to the average inter-particle distance and used during visualisation in Section 5. The number of particles, and their displacement fromthe galactic plane, are defined by N star and σ star respec-tively and determine the thickness of the disk. These canbe calculated using one of three available models:1. A Gaussian thickness defined by tunable input pa-rameter σ xyz , where particles are distributed with σ star = σ xyz and N star is scaled with I s . As such,brighter pixels of the image are represented by morepoints across a larger spatial volume. The scaling of N star with I s is defined as the inverse of the functiondefining I s from I p , such that N star · I s ≈ I p .2. Radial ( H r ) and axial ( H z ) parameters are provided,along with tunable input parameter σ xy . N star isscaled proportionally to exp ( − R/H z ) (where R isradial distance of the particle to the center of thedisk), and the particles are distributed with σ xy inthe radial plane and σ z axially, representing a ta-pered disk structure.3. Thickness is scaled to fit to known measurements offlared disk thickness in edge-on galaxies, followingthe measurement of F W HM z,g (gas layer thickness)as presented in Figure 25 of [38].In general, the appropriate model is chosen to matchobserved properties. Model 1 may be used as a heuristicapproach approximating a Gaussian thickness. Model 2provides a disk thickness matching generally observed ta-pered galactic disks, whilst Model 3 provides a thicknessmatching observations of flared H i disk galaxies. In thecase of M83, Model 2 is used for the stellar disk compo-nents, whilst Model 3 is used for the flared gaseous H i disk.The model may be chosen per-component in the parame-ter file passed to the galaxy modeller stage of the pipelineshown in Figure 2.As can be seen in high quality edge-views of spiralgalaxies, for example that shown in Figure 1 (right), dustlanes can have complex morphologies with cloud and filament-like structures, and can be very well-defined against thebright stellar background. To reflect this, the thickness ofthe dust component is implemented using a slightly dif-ferent scheme. The point clouds generated around seedparticles are distributed using one of two models:1. Dust filaments are approximated via assigning ve-locities to particles randomly distributed around thelinear velocity defined by the TiRiFiC model. A ran-dom walk is traced, generating a new particle at eachstep, with a gravitational factor applied such that fil-aments are constrained near to the galactic plane.2. As with the previous item, however particle velocitiesare distributed around the rotational velocity of thedisk, constraining filaments to the same rotation asthe disk.As discussed further in Section 7, the dust thicknessmodel is addressed with a more heuristic approach thanthe disk, and so the appropriate model is chosen for visualeffect.7he galactic bulge is more simply defined as a Gaussiandistribution of points defined by input parameters σ x , σ y , σ z , and N bulge , and coloured white ( RGB bulge = [1 , , N p , N star , N bulge ) are provided as user inputs, whilst the final num-ber of particles depends on the input image (a more sat-urated image results in production of more particles, asthere is more emission to emulate). Currently these initialvalues are tuned empirically based on visualisation of theresulting model using an Earth-based view and compari-son to observations; too many particles result in an over-saturated model where individual features are difficult todistinguish, whilst too few result in an under-saturatedmodel where features are missing or not well resolved.All the components described above can be combinedto create complex galaxies, like M 83. However, each differ-ent component can also be used to model simpler objects,like elliptical or dwarf galaxies. The former can be repre-sented as a stellar ellipsoid, or nested ellipsoidal distribu-tions representing a bulge and extended star distribution,characterised by different values for axis and star density.The latter may represent various types of small galaxiesthat can have either a disk-like structure, in which casethe stellar population model is used, or a spherical shape,in which case the galactic bulge model is used. For bothtypes of small galaxy, a hydrogen gas encompassing cloudcan be present, which may be added using the diffuse gas component. We have used these models in Section 6 toconstruct and visualise the M 83 local group of galaxies.Finally, cleaned and pre-processed observational im-ages (as described in Section 4.2) together with the binaryimage masks are used to provide RGB s colour values topaint the galaxy population. In case where images arenot available (typically the case with bulges or globularclusters) a colour can be chosen manually to visually dis-tinguish each of the galactic components, and if possible toresemble existing recognisable false-colour composite im-ages. Each of the components is generated sequentially,and the full galaxy representation is then composited andwritten as one of a series of commonly available file typesin astronomy (CSV, Gadget, or HDF5). All of the param-eters can be provided though an input key-value text fileshown as create galaxy.par in Figure 2. Once several galaxies have been modelled, they can becombined within the same scene to render, for example,a group of neighbouring objects bound by gravity in thesame system. This could represent a local group of galaxies made of several components, like the Milky Way and theMagellanic Clouds, or a cluster of galaxies, like the Comaor the Virgo clusters, composed of hundreds to thousandsof galaxies.The various components are placed in the position iden-tified by their known astronomical coordinates, convertedto a Cartesian reference frame for compatibility with typi-cal visualisation tools. Their size may be arbitrarily scaledfor visibility; the distance between galaxies is typicallyorders of magnitude bigger than their size, such scalingis then required for concurrent visualisation of multipleobjects with a moving camera. The scaling may be thesame for all elements in the scene, to preserve the sizeratio between objects, or specific to each member to high-light given elements of the group. In both cases, how-ever, the relative positions and orientations of galaxies arepreserved, such that the resulting image forms a realistic,although locally magnified, representation of the actualsystem. Galaxy combinations are merged after construc-tion, creating a single dataset that may be visualised. Anapplication of such combination is presented in Section 6,describing the M 83 galaxy local group.
5. 3D Galaxy Visualisation
This section describes the visualisation process for aconstructed galaxy model. We utilise the astronomicalvisualisation tool Splotch, which is well-suited to our pur-pose as it is designed for particle-based astronomy dataand supports multiple species of particle, natively support-ing the multi-component particle-based structure of ourgalaxy models. The software, with which the authors havesignificant development experience, is open-source whichallowed us to modify the underlying rendering algorithm tobetter support our galaxy models as described in Section5.1. Future work (as discussed in Section 8) will also ben-efit from the high-performance parallel nature of Splotchas we scale up from single galaxies to large groups.
Splotch is a high performance and scalable scientific vi-sualisation package designed for particle-based astronom-ical datasets, with implementations for a variety of hard-ware platforms [11, 23, 47, 12]. Splotch has been utilisedin the past for a variety of types of visualisation, includingillustrating numerical simulation results in academic talks,scientific communication and outreach through generationof movies for planetariums, and supplementing scientificanalysis through visualisation in the context of theoreticalvirtual observatories [46, 13].The Splotch software is written in C++, with min-imal dependencies beyond those for parallel models (e.g.OpenMP , CUDA , the Message Passing Interface (MPI) ) https://developer.nvidia.com/cuda-zone B -Spline[34], defined such that particles overlap with a set numberof neighbours. Viewer ImageSpace ObjectSpace
Figure 6: Volume splatting for particle data: particles are projectedfrom object space to image space, and splatted across the image usinga footprint function, typically a Gaussian kernel.
The rendering method of Splotch is an implementa-tion of volume splatting [57]. First each data element istransformed relative to a viewpoint, a parallel or perspec-tive projection applied, and then the contribution of eachelement to line-of-sight rays cast from image pixels is com-puted using a “splatting” kernel, summarised in Figure 6.In Splotch, each data element is represented by a particle,and a simplified emission and absorption optical model([31]) is used to define each particle’s contribution to therays as follows, starting from the radiative transfer equa-tion in differential form: dI ( x ) dx = ( E p − A p I ( x )) ρ p ( x ) (1) x is the coordinate along the line of sight, I is the intensityat position x , E p and A p are the emission and absorptioncoefficients of particle p . In this form, both E p and A p directly rely on ρ p ( x ), which defines the contribution tomatter density interpolated from particle p , and is definedusing a Gaussian distribution: ρ p ( x ) = ρ ,p exp( − || x − x p || /σ p ) (2)with x p representing the particle coordinates, and ρ ,p and σ p being the mass density and the radius of the particlerespectively. For a more convenient compact support, thedistribution is truncated at χ · σ p , where χ is a suitablydefined factor typically chosen such that χ · σ p ≈ h ; where h is the intrinsic smoothing length of the particles (e.g.as described in [25]). Due to this relation, the particleradius is commonly referred to as the smoothing lengthin this document and the referenced Splotch publications.As noted in [11], the B -Spline typical for SPH particlesis very similar in shape to the Gaussian distribution usedhere. Following on from Equations 1 and 2, the contributionof a single particle to a ray is defined as: I after = ( I before − E p /A p ) exp( − A p ∞ (cid:90) −∞ ρ p ( x ) dx ) + E p /A p (3)For simplicity, the frequency dependency of the inten-sity is not included in Equations 1 to 3; however, such adependency does exist. The transfer function of Splotchsupports emission, and absorption as a function of emis-sion, in three frequencies corresponding to colors R, G andB, and referred to hereafter as floating point triplets E RGB and A RGB respectively. Furthermore, each particle has anintrinsic type property, used to distinguish particle species(for example gas, stars, and black holes). The transferfunction can additionally be customised per type, to sup-port an arbitrarily large range of functions each with threefrequency outputs.This optical model requires the particle data to besorted back-to-front with respect to the viewer, such thatabsorption and emission can be integrated in the correctorder. Splotch also supports a further simplification, anassumption that E p = A p removes the need for order de-pendent rendering, and acts as a high performance approx-imation for highly diffuse or optically thin material (e.g.intergalactic medium), or extremely compact and brightmaterial (e.g. stars), both of which are common in astro-physical simulation. As such, a flag can be set to indicate E p = A p , neglecting sorting before rendering, however thismode is not used for the model visualisation in Section 5.2due to the described treatment of stellar dust.We extended the Splotch rendering software to treatemission and absorption coefficients independently, ratherthan treating absorption as a function of emission, to sup-port an absorptive galactic dust component. The extendedSplotch code allows the user to provide per particle coeffi-cients for both emission and absorption, in three frequen-cies, from their source data. These are included in theparallelised radiative transfer computation in the render-ing kernel at a cost of 3 additional floating point fields perparticle, or a 30% increase in memory consumption duringexecution. To support this independent absorption coeffi-cient, an algorithmic extension is also needed to accountfor the lower limit of absorption, i.e. for a purely emit-ting particle or one with negligible absorption, A p ≈ I after = I before + E p ∞ (cid:90) −∞ ρ p ( x ) dx (4)The additional fields and extended rendering algorithmare implemented with C pre-processor commands allowingthem to be switched off at compile time. This supports aquick reversion to the simplified rendering model, in the9 alactic Component Source Imaging Data Stellar distribution (Optical) SINGG H α - and R -bandStellar distribution (UV) GALEX Near and Far UVDiffuse hydrogen gas ATCA H i at three resolutionsDust Spitzer IRAC 8 µ mGalactic Bulge None Table 2: A mapping demonstrating the relationship of galactic compo-nents used for M 83 to source image data. case where there is no sensible means of modelling theabsorption.
Splotch takes as input: the data file written by theGalaxy Modeller (e.g.
NgcXXXX 0000 as shown in Fig-ure 2), a key-value parameter file ( visualise galaxy.par ) de-scribing the scene configuration, and optionally a scene file( path.scene ) which can be used to describe a set of sceneconfigurations for a movie.For each of the galactic components, which are treatedas separate particle species in Splotch, a series of tunablevisual parameters are available. The smoothing length h of particles can be scaled using a size parameter, and theintensity I can also be scaled using a brightness parameter.The emission and absorption coefficients of each com-ponent are defined by the intensity I of the particle, re-tained as I s from Section 4.3, such that the emissivity ofthe particles is directly related to the observed intensitiesof the galactic component to which the particle belongs.The emission coefficient in each frequency of the final im-age E RGB is defined as I s · RGB s . The absorption coef-ficient A RGB is defined as I s · RGB A , where RGB A is athree component absorption profile provided as a lookuptable during transfer function evaluation.In the example case for M 83, there are 5 galactic com-ponents included in the visualisation, which are informedby observations as illustrated by Table 2. The stellar dis-tribution and diffuse hydrogen gas are treated as colouredemissive sources, where E RGB = I s · RGB s and A RGB = 0.The bulge component is treated as a white fully emissivesource, where E RGB = I s · [1 , ,
1] and A RGB = 0. Thedust is defined as a grey fully absorptive source, where
RGB s = [0 , ,
0] and A RGB = I s · [1 , ,
6. Reconstruction and Visualisation of M 83 andits local group
Figure 7 demonstrates the results of applying our re-construction and visualisation methodology to the M 83galaxy. The images are split into far and close ( left and right respectively), showing face-on, angled, and edge-onviews ( top to bottom respectively). The far images includethe H i extended gaseous disk, illustrating the large warpedstructure which is not easily discernible for an astronomerviewing the observed H i images (Figure 4). The maxi-mum extent of the H i region is ∼
100 kilo-parsecs (kpc) [26]. In contrast the close images show the inner stellardisk, which is ∼
13 kpc diameter [53], and have the H i re-moved to more closely resemble an optical image (such asFigure 1 left ). These close images highlight the spiral armstructure, absorptive dust lanes, tapered disk and stellarbulge.The reconstruction and visualisation was performed us-ing a single dual-socket node of a Cray XC50, with two 22-core Intel Broadwell processors clocked at 2.2 Ghz, and 128GB of DDR4-2400 memory. The generated M 83 modelconsists of approx. 22 million particles, split amongst thecomponents as shown in Figure 8, which depicts the per-component computational time required for reading sourceimage files, reconstruction, and colouring, totalling ≈ i , two for UV, for example).Reconstruction is not directly dependent on the numberof particles used for the component, as the particle countfor the constructed component is a result of the saturationof the input image and the parameters determining seedparticle count and surrounding point clouds, as discussedin Section 4.The overall cost in terms of computational time for asingle galaxy is, in general, acceptable; however, if thismethodology were extended to model a large combinedgroup with, for example, 10 to 1000 galaxies, the cost isexpected to increase linearly, growing unreasonably highfor 1000 galaxies. In this case, it is expected that non-interacting galaxies could be modelled independently toallow trivial parallelisation on a per-galaxy basis (e.g. onegalaxy per-node of a compute cluster), whilst galaxy com-ponents could also be constructed in parallel to improvethe per-galaxy computational cost (e.g. one or more coresof a compute node per galaxy component). Similarly,Figure 9 shows that memory consumption can extend toapproximately 5-6 times the size of the generated modelthroughout the program lifetime, which is acceptable fora single galaxy but a large group of combined galaxiesprocessed in parallel may require further efforts to reducethe memory footprint. The plateaus seen in Figure 9 arecaused by allocating a large block of memory for particledata, and reusing this block throughout the reconstructionprocess for different components.Visualisation is performed using the Splotch softwarebuilt with OpenMP support, and run with 44 OpenMPthreads to match the number of cores available on the testplatform. Figure 10 shows the visualisation time for eachof the sub-figures of figure 7, split into the key componentsof the visualisation process. Transformation and colouringof particles are highly parallel operations taking ≈ . s in-dependent of the scene. Sorting is more computationallyintensive, requiring ≈ s per frame, also inherently in-dependent of the scene configuration. Rendering is scene-dependent, ranging from 1-3 seconds per scene for the stel-10a) (d)(b) (e)(c) (f) Figure 7: Example visualisation outputs for the M 83 galaxy, split into far and close ( left and right , respectively), showing face-on, angled,and edge-on views ( top to bottom , respectively). The far images (a, b, c) include the extended gaseous disk ( ∼
100 kpc diameter), illustratingthe large warped structure. The close images (d, e, f) zoom into the stellar disk and have the H i removed to more closely resemble an opticalcomposition ( ∼
17 kpc diameter), and show clearly the absorptive dust lanes, tapered disk and stellar bulge. The added image scales give asmaller impression, as the edges of the galaxy fade into darkness, implying a slightly smaller structure. i disk is visible. This scene dependence stems from the av-erage radius of particles, as discussed in [47][12], a largerparticle radius affects more pixels in the output image,which reduces parallel rendering performance. From theseresults, we can expect that during an extended movie ren-dering the gaseous H i scenes will take approximately 10xthe time of stellar scenes, as larger particles are utilised tocreate the gaseous effect seen in the H i scene. Optical Stars10.2M UV Stars1.6M HI gas8.1M Dust2.6M Bulge100K T i m e / s Galaxy Component
ReconstructionColouringImage reading
Figure 8: Performance of the galaxy reconstruction code for the M 83example case. The number of particles for each component is listedunder the horizontal axis component labels. . . . . . . . . . . . . . . . . . . . . . . . . . . M e m o r y U t ili s a t i o n / M B Program Lifetime (% of instructions executed)
Size of reconstructed 3D model on disk
Figure 9: Memory consumption during program lifetime for thegalaxy reconstruction code for the M 83 example case.
Figure 11 demonstrates another view of M 83, as amember of a group of galaxies known as the M 83 localgroup. The top four panels show M 83 as the most promi-nent object of group, alongside neighbouring dwarf galax-ies, from a variety of viewing angles. According to obser-vations, these galaxies may have either a disk-like shapeor an almost spherical symmetry, and they have been re-constructed by our methodology following the proceduredescribed in Section 4.3 for dwarf galaxies. The sourcedata for these objects is not high resolution, but sufficient (a) (b) (c) (d) (e) (f) T i m e / s Figure
RenderSortColourTransform
Figure 10: Rendering performance of the Splotch code for the M 83galaxy images seen in subfigures (a,b,c,d,e,f) of Figure 7. The galaxymodel consists of approximately 22 million particles. to give an indication of their main features and distinguishbetween disk or elliptical galaxies for visualisation both inoptical and in H i (no data is available for possible dustdistributions). The top-left panel shows the galaxy groupfrom an Earth-based observing direction, centred on M 83(the brightest object), and including the star distributionof the four closest members of the local group. In the op-tical band the galaxies are just visible dots. However, assoon as H i is added (top right panel), a much richer sceneappears from the same point of view, with M 83 show-ing the complex H i distribution already highlighted aboveand the four dwarfs all clearly visible, with the gas distri-bution much more extended than the stellar distribution.The bottom left panel show a zoom-in of two of the mem-bers of the local group, a disk galaxy NGC5264 on the left,and IC4316 on the right. The bottom right panel showsa further zoom in of the dwarf galaxy IC4316, showinga more detailed view of the mixed star distribution andsurrounding H i cloud.The two middle panels of Figure 11 show the galaxygroup from two different points of view, neither of whichare Earth-based, demonstrating a unique view of these ob-jects that is only possible with a reconstruction and visu-alisation methodology such as ours. These novel viewshighlight that the visual impression of galaxies shape, po-sition, and structure, relative to one another can be mis-leading, due to projection effects of a 2D-only view. Fur-thermore, these different representations can provide hintsto astronomers regarding features of the system, such asevolutionary interaction between neighbours. One possibleexplanation for the gas structure of M 83 could be that pastinteractions with other galaxies in the local group causethe observed tilted and elongated tails which are not visi-ble in the star distribution. Viewing the middle left panelcould suggest that the upper gas tail of M 83 may relateto a past interaction with the dwarf immediately above it.However, from the alternate point of view presented in the12iddle right panel, we observe that the tail is no longerpointing to the same dwarf, which moved to the top rightcorner of the image. Hence, a direct influence of the dwarfon M 83’s gas distribution appears unlikely, directing thescientist toward alternative explanations.
7. Physical Realism
Our presented methodology aims to improve the anal-ysis of the structure of galaxies based on combining kine-matic and image information for observed galaxies. Ingeneral, it is not feasible to observe galaxies from an an-gle other than that of an Earth-based observer, and assuch impossible to be certain that our model is struc-turally accurate for the observed galaxy. One approachfor validation is to compare to observations of galaxiesseen from Earth at similar viewpoints. Figure 12 comparesa sample of our generated M83 images from non-Earthbased viewpoints to similar shaped galaxy views extractedfrom observational data, highlighting the capability of ourmethodology to achieve a realistic result based on similarobservations of other galaxies.We aim to build this physical realism into the galaxyrepresentation using state of the art multi-wavelength ob-servational data complemented by numerical representa-tions of analytical models for different components of galax-ies as inferred from observations. To evaluate the extentto which this is achieved during construction, a series ofhigh level morphological and visual properties have beenidentified with an aim to comprehensively describe thereconstruction process. The methods used to determinethese properties for each galaxy component have been cat-egorised by the extent to which physically realistic meth-ods are utilised. The properties, with reference to theprevious sections, are:
Radial seed distribution
This property defines the dis-tribution of seed particles (as discussed in Section4.3) in the radial plane of the galaxy (i.e. distribu-tion across the galaxy face).
Axial seed distribution
This property defines the dis-tribution of seed particles in the axial plane of thegalaxy (i.e. galaxy thickness).
Particle size and distribution
This property defines thesize of each representative particle and distributionof such particles around the seed (i.e. galaxy den-sity).
Emission
This property defines the value used as theemission coefficient (as discussed in Section 5.1) foreach particle during visualisation (i.e. galaxy bright-ness and colour).
Absorption
This property defines the value used as theabsorption coefficient for each particle during visual-isation (ie. galaxy absorption, dust lanes, and shad-ows). The method or model used to define each of these prop-erties per galaxy component has been categorised into oneof the following groups, ordered by physically realism, withthe top being most realistic:
Source data
The property has been derived directly fromobservational data of the source object being mod-elled.
Observed model
The property has been derived fromgeneral observations or relations observed for the ob-ject, or type of object, being modelled.
Theoretical model
The property has been derived froma theoretical model or simulation of the object, ortype of object, being modelled.
Scaled to source data
The property has been scaled ac-cording to source data, but may be initially definedby heuristic.
Heuristic model
The property has been defined and tunedheuristically in collaboration with an experienced as-tronomer based on resultant visual effects.As illustrated in Table 3, whilst much of the mod-elling and visualisation methodology exploits observed orinferred data, there are a series of heuristics requiring man-ual intervention which introduces a bias in our results. Inparticular, the initial number of particles per component iscurrently empirically tuned based on visualisation results,we envision a more robust approach to determining ini-tial particle counts considering observed parameters (e.g.estimated mass of each component) for the galaxy beingmodelled. Further work is envisioned to develop a morerealistic axial dust model, considering the existing researchinto the morphology and scattering effects of dust grainsin both in astronomy and visualisation research such as[30], who use a 3D Perlin noise model [42] to describe dustmorphology combined with careful treatment of dust scat-tering, or [35] who similarly use Perlin noise for a morerealistic gaseous nebulae effect. Furthermore, the axialdistribution of the bulge component may be more robustlydefined based on observed properties in systematic studiesof bulges in galaxies, for example [16, 15].
8. Conclusions and Future Work
We have presented a novel 3D modelling and visualisa-tion methodology for the reconstruction of nearby galax-ies based on a wide variety of multi-resolution observedimages and derived data. This includes optical, UV, IRand H i images along with H i kinematic models of nearbygalaxies based on tilted-ring fitting. We addressed severalobjectives: (1) to create realistic views of galaxies fromviewpoints that are not otherwise possible to observe from,(2) explore the validity of derived spatial 3D models of suchgalaxies, and (3) allow enhanced visual analysis of the 3D13 C4316 IC4247NGC5264NGC5423
ESO444NGC5264 IC4316
ESO444 IC4316NGC5264
NGC5264 IC4316
IC4316
Figure 11: The M 83 galaxy and its neighbouring dwarf galaxies. Top left, the M 83 group in the optical band. Top right, the M 83 groupwith H i emission added. Middle panels, the M 83 group from different observation angles. Bottom left, two of the dwarfs companions ofM 83. Bottom right, a close up of dwarf companion IC4316. As this is a 3D rendering with multiple objects at different distances, we do notinclude a linear scaling in these figures. alaxyComponent Component PropertyRadial seeddistribution Axial seeddistribution Particle sizeand distribution Emission Absorption Bulge Observed model Heuristic model Heuristic model Heuristic model NoneStellar Source data Observed model Heuristic model Scaled tosource data NoneDust Source data Heuristic model Heuristic model Scaled tosource data Scaled tosource dataDiffuse gas Source data Theoretical model Heuristic model Scaled tosource data None
Table 3: A summary of the underlying models used for the various configurable properties of each galaxy component. galaxy morphology (stars, gas and dust). Our method-ology allows dynamic 3D exploration of galaxies and canprovide new views (see Figure 7) of galaxies that are onlytypically observable from an Earth-based viewpoint. Suchnovel visualisations are able to illustrate several differentgalaxy components, e.g., extended H i disks and inner stel-lar disks, only the latter of which is typically well knownto the public, and show the 3D structure of the extendedH i disk of M83, incorporating kinematic models in our re-construction to represent the warped structure that is nottypically visible in 2D observed images.We further enable comparison of 3D reconstructed galax-ies with other observed galaxies at different viewing an-gles (as in Figure 12), which could support evaluation ofthe kinematical models that inform their 3D shapes. Ourmethodology allows enhanced analysis of these galaxies,e.g. allowing the viewer to distinguish whether observedfeatures are real characteristics or projection effects, orhighlight properties hidden by a 2D representation, asdemonstrated in Figure 11. We note that galaxies ob-served close to edge-on can be difficult to reconstruct in3D as detailed information of their spiral structure is noteasily available, as such the current methodology is notable to present a realistic face-on view of observed edge-ongalaxies. One approach to address this problem from a vi-sual perspective could be to generate a believable face-onimpression through statistical analysis of existing obser-vations of face-on galaxies, or exploiting recent results ofhigh resolution cosmological simulations such as EAGLE[32] to realistically describe the spiral arms in galaxies.Work is on-going focusing on introducing more auto-mated pipeline stages, enabling new galaxies to be mod-elled faster and reducing the bias introduced by manual in-tervention as discussed in Section 7. The existing pipelinefor both the modelling and visualisation is command-linebased, with key-value parameter files as inputs; to im-prove uptake and streamline the modelling and visuali-sation process we are considering the viability of a moreuser-friendly interface such as a Python scripting layer orgraphical interface. We then intend to release the mod-elling code, alongside full instructions for source data col- lection and preparation, alongside the next major releaseof the Splotch visualisation code , for use on Linux andMacOS systems.In future, we plan to enable interactive visualisationsand fly-throughs of the 3D reconstructed galaxies and galaxygroups. Such fly-through movies, or even interactive ex-ploration of single galaxies or galaxy groups, are envisagedto provide new opportunities for visual discovery such asthose discussed in Section 6. We believe it may also beinteresting to explore the combination of our modellingapproach with more cinematic rendering tools, such asutilised in the works of [36, 6], relaxing the requirementson physical realism to improve visual impact for outreachpurposes.Finally, our galaxy group demonstration is the firststep toward a comprehensive methodology able to providea visual representation of large galaxy assemblies based ontheir observational properties. A survey-like approach willrequire streamlining of the manual, user-defined, stages ofour pipeline, which would quickly become the most lim-iting scaling factor, however holds significant analysis op-portunities and could, for example, lead to a generallyaccessible dataset holding many 3D referential images ofa variety of galaxies. We further hope that in future thismay lead to more advanced analysis scenarios of dynamicscenes, for example exploring a dynamic scene of mov-ing galaxy neighbours that demonstrates possible past orfuture states of a group based on observed inter-galaxykinematics. Acknowledgements
KD acknowledges support by the ORIGINS cluster,funded by the Deutsche Forschungsgemeinschaft (DFG,German Research Foundation) under Germany’s Excel-lence Strategy, EXC-2094, 390783311. MK acknowledgessupport by NEANIAS, funded by the EC Horizon 2020 re-search and innovation programme under grant agreementNo. 863448. https://github.com/splotchviz/splotch igure 12: A comparison of our M83 visualisations from non-Earth-based viewpoints to similar observations of other galaxies from Earth.The observed images (bottom row) have been rotated and scaled to match our images for comparison. On the left, we compare ourvisualisation of the M83 dusty stellar disk ( top left ) to NGC7331, a two-color composite image ( bottom left ) from the Digitised Sky Survey(see acknowledgements) extracted using Aladin Lite [5]. In the center, we compare our visualisation of the edge-on M83 stellar disk ( topcenter ) to NGC 4565 ( bottom center ), acknowledged as in Figure 1. On the right we compare our visualisation of the warped M83 H i gasdisk ( top right ) with the warped hydrogen disk of UGC 3697 ( bottom right ), courtesy of NRAO/AUI/NSF. Figure 12 is based on photographic data of the Na-tional Geographic Society Palomar Observatory Sky Sur-vey (NGS-POSS) obtained using the Oschin Telescope onPalomar Mountain. The NGS-POSS was funded by agrant from the National Geographic Society to the Cal-ifornia Institute of Technology. The plates were processedinto the present compressed digital form with their per-mission. The Digitized Sky Survey was produced at theSpace Telescope Science Institute under US Governmentgrant NAG W-2166.
References [1] Bandyopadhyay, R., Das, R., Mondal, S., Ghosh, S., 2020. Mor-phology and ionization characteristics of planetary nebulae PB1 and PC 19. MNRAS 496, 814–831.[2] Bertin, E., 2009. SkyMaker: astronomical image simulationsmade easy. Mem. Soc. Astron. Italiana 80, 422.[3] Bertin, E., Arnouts, S., 1996. SExtractor: Software for sourceextraction. A&AS 117, 393–404.[4] Biffi, V., Dolag, K., B¨ohringer, H., Lemson, G., 2012. Observingsimulated galaxy clusters with PHOX: a novel X-ray photonsimulator. MNRAS 420, 3545–3556.[5] Boch, T., Fernique, P., 2014. Aladin Lite: Embed your Sky inthe Browser, in: Manset, N., Forshay, P. (Eds.), AstronomicalData Analysis Software and Systems XXIII, p. 277. [6] Borkiewicz, K., Naiman, J.P., Lai, H., 2019. Cinematic Vi-sualization of Multiresolution Data: Ytini for Adaptive MeshRefinement in Houdini. AJ 158, 10.[7] Bottrell, C., Torrey, P., Simard, L., Ellison, S.L., 2017. Galax-ies in the Illustris simulation as seen by the Sloan Digital SkySurvey – I. Bulge+disc decompositions, methods and biases.MNRAS 467, 1033–1066.[8] Crocce, M., Castander, F.J., Gazta˜naga, E., Fosalba, P., Car-retero, J., 2015. The MICE Grand Challenge lightcone simula-tion - II. Halo and galaxy catalogues. MNRAS 453, 1513–1530.[9] Dale, D.A., Cohen, S.A., Johnson, L.C., Schuster, M.D.,Calzetti, D., Engelbracht, C.W., Gil de Paz, A., Kennicutt,R.C., Lee, J.C., Begum, A., Block, M., Dalcanton, J.J., Funes,J.G., Gordon, K.D., Johnson, B.D., Marble, A.R., Sakai, S.,Skillman, E.D., van Zee, L., Walter, F., Weisz, D.R., Williams,B., Wu, S.Y., Wu, Y., 2009. The Spitzer Local Volume Legacy:Survey description and infrared photometry. ApJ 703, 517–556.[10] Dolag, K., Borgani, S., Murante, G., Springel, V., 2009. Sub-structures in hydrodynamical cluster simulations. MNRAS 399,497–514.[11] Dolag, K., Reinecke, M., Gheller, C., Imboden, S., 2008.Splotch: visualizing cosmological simulations. New Journal ofPhysics 10, 125006.[12] Dykes, T., Gheller, C., Rivi, M., Krokos, M., 2017. Splotch:porting and optimizing for the xeon phi. The InternationalJournal of High Performance Computing Applications 31, 550–563.[13] Dykes, T., Hassan, A., Gheller, C., Croton, D., Krokos, M.,2018. Interactive 3D visualization for theoretical virtual obser-vatories. MNRAS 477, 1495–1507.
14] Fedorov, A., Beichel, R., Kalpathy-Cramer, J., Finet, J.,Fillion-Robin, J.C., Pujol, S., Bauer, C., Jennings, D., Fennessy,F., Sonka, M., Buatti, J., Aylward, S., Miller, J.V., Pieper, S.,Kikinis, R., 2012. 3D Slicer as an image computing platformfor the Quantitative Imaging Network. Magnetic ResonanceImaging 30, 1323 – 1341. Quantitative Imaging in Cancer.[15] Fisher, D.B., Drory, N., 2011. Demographics of Bulge typeswithin 11 Mpc and implications for galaxy evolution. ApJ 733,L47.[16] Gadotti, D.A., 2009. Structural properties of pseudo-bulges,classical bulges and elliptical galaxies: a Sloan Digital Sky Sur-vey perspective. MNRAS 393, 1531–1552.[17] Garc´ıa-D´ıaz, M.T., L´opez, J.A., Steffen, W., Richer, M.G.,2012. A detailed morpho-kinematic model of the Eskimo, NGC2392: A unifying view with the Cats Eye and Saturn planetarynebulae. ApJ 761, 172.[18] Gil de Paz, A., Boissier, S., Madore, B.F., Seibert, M., Joe,Y.H., Boselli, A., Wyder, T.K., Thilker, D., Bianchi, L., Rey,S.C., Rich, R.M., Barlow, T.A., Conrow, T., Forster, K., Fried-man, P.G., Martin, D.C., Morrissey, P., Neff, S.G., Schimi-novich, D., Small, T., Donas, J., Heckman, T.M., Lee, Y.W.,Milliard, B., Szalay, A.S., Yi, S., 2007. The GALEX UltravioletAtlas of Nearby Galaxies. ApJS 173, 185–255.[19] Guedes, J., Callegari, S., Madau, P., Mayer, L., 2011. Form-ing realistic late-type spirals in a ΛCDM Universe: The ErisSimulation. ApJ 742, 76.[20] Hassan, A., Fluke, C.J., 2011. Scientific visualization in astron-omy: Towards the petascale astronomy era. PASP 28, 150–170.[21] Helou, G., Roussel, H., Appleton, P., Frayer, D., Stolovy, S.,Storrie-Lombardi, L., Hurt, R., Lowrance, P., Makovoz, D.,Masci, F., Surace, J., Gordon, K.D., Alonso-Herrero, A., Engel-bracht, C.W., Misselt, K., Rieke, G., Rieke, M., Willner, S.P.,Pahre, M., Ashby, M.L.N., Fazio, G.G., Smith, H.A., 2004. Theanatomy of star formation in NGC 300. ApJS 154, 253–258.[22] Hildebrand, K., Magnor, M., Fr¨ohlich, B., 2006. 3D reconstruc-tion and visualization of spiral galaxies. WSCG 14, 113–120.[23] Jin, Z., Krokos, M., Rivi, M., Gheller, C., Dolag, K., Reinecke,M., 2010. High-performance astrophysical visualization usingsplotch. Procedia Computer Science 1, 1775–1784.[24] J´ozsa, G.I.G., Kenn, F., Klein, U., Oosterloo, T.A., 2007. Kine-matic modelling of disk galaxies. I. A new method to fit tiltedrings to data cubes. A&A 468, 731–774.[25] Koda, J., Sofue, Y., Wada, K., 1999. How to determine thesmoothing length in sph?, in: Miyama, S.M., Tomisaka, K.,Hanawa, T. (Eds.), Numerical Astrophysics, Springer Nether-lands, Dordrecht. pp. 391–392.[26] Koribalski, B.S., Wang, J., Kamphuis, P., Westmeier, T.,Staveley-Smith, L., Oh, S.H., L´opez-S´anchez, ´A.R., Wong, O.I.,Ott, J., de Blok, W.J.G., Shao, L., 2018. The Local Volume HI Survey (LVHIS). MNRAS 478, 1611–1648.[27] Lee, K.G., Krolewski, A., White, M., Schlegel, D., Nugent,P.E., Hennawi, J.F., M¨uller, T., Pan, R., Prochaska, J.X., Font-Ribera, A., Suzuki, N., Glazebrook, K., Kacprzak, G.G., Kar-taltepe, J.S., Koekemoer, A.M., F`evre, O.L., Lemaux, B.C.,Maier, C., Nanayakkara, T., Rich, R.M., Sanders, D.B., Sal-vato, M., Tasca, L., Tran, K.V.H., 2018. First data release ofthe COSMOS ly α mapping and tomography observations: 3dly α forest tomography at 2.05 < z < .[55] Vohl, D., Fluke, C.J., Barnes, D.G., Hassan, A.H., 2017. Real-time colouring and filtering with graphics shaders. MNRAS 471,3323–3346.[56] Wen, Z., O’dell, C.R., 1995. A three-dimensional model of theOrion Nebula. ApJ 438, 784–793.[57] Westover, L.A., 1991. Splatting: A parallel, feed-forward vol-ume rendering algorithm. Ph.D. thesis. Chapel Hill, NC, USA.UMI Order No. GAX92-08005.[58] York, D.G., Adelman, J., Anderson, John E., J., Anderson,S.F., Annis, J., Bahcall, N.A., Bakken, J.A., Barkhouser, R.,Bastian, S., Berman, E., Boroski, W.N., Bracker, S., Briegel, C.,Briggs, J.W., Brinkmann, J., Brunner, R., Burles, S., Carey, L.,Carr, M.A., Castander, F.J., Chen, B., Colestock, P.L., Con-nolly, A.J., Crocker, J.H., Csabai, I., Czarapata, P.C., Davis,J.E., Doi, M., Dombeck, T., Eisenstein, D., Ellman, N., Elms,B.R., Evans, M.L., Fan, X., Federwitz, G.R., Fiscelli, L., Fried-man, S., Frieman, J.A., Fukugita, M., Gillespie, B., Gunn, J.E.,Gurbani, V.K., de Haas, E., Haldeman, M., Harris, F.H., Hayes,J., Heckman, T.M., Hennessy, G.S., Hindsley, R.B., Holm, S.,Holmgren, D.J., Huang, C.h., Hull, C., Husby, D., Ichikawa,S.I., Ichikawa, T., Ivezi´c, ˇZ., Kent, S., Kim, R.S.J., Kinney,E., Klaene, M., Kleinman, A.N., Kleinman, S., Knapp, G.R.,Korienek, J., Kron, R.G., Kunszt, P.Z., Lamb, D.Q., Lee, B.,Leger, R.F., Limmongkol, S., Lindenmeyer, C., Long, D.C.,Loomis, C., Loveday, J., Lucinio, R., Lupton, R.H., MacKin-non, B., Mannery, E.J., Mantsch, P.M., Margon, B., McGehee,P., McKay, T.A., Meiksin, A., Merelli, A., Monet, D.G., Munn,J.A., Narayanan, V.K., Nash, T., Neilsen, E., Neswold, R., New-berg, H.J., Nichol, R.C., Nicinski, T., Nonino, M., Okada, N.,Okamura, S., Ostriker, J.P., Owen, R., Pauls, A.G., Peoples,J., Peterson, R.L., Petravick, D., Pier, J.R., Pope, A., Pordes,R., Prosapio, A., Rechenmacher, R., Quinn, T.R., Richards,G.T., Richmond, M.W., Rivetta, C.H., Rockosi, C.M., Ruth-mansdorfer, K., Sand ford, D., Schlegel, D.J., Schneider, D.P.,Sekiguchi, M., Sergey, G., Shimasaku, K., Siegmund, W.A.,Smee, S., Smith, J.A., Snedden, S., Stone, R., Stoughton, C.,Strauss, M.A., Stubbs, C., SubbaRao, M., Szalay, A.S., Sza-pudi, I., Szokoly, G.P., Thakar, A.R., Tremonti, C., Tucker,D.L., Uomoto, A., Vanden Berk, D., Vogeley, M.S., Waddell,P., Wang, S.i., Watanabe, M., Weinberg, D.H., Yanny, B., Ya-suda, N., SDSS Collaboration, 2000. The Sloan Digital SkySurvey: Technical Summary. AJ 120, 1579–1587.[59] ZuHone, J.A., Kowalik, K., Ahman, E., Lau, E., Nagai, D.,2018. The galaxy cluster merger catalog: An online repositoryof mock observations from simulated galaxy cluster mergers.ApJS 234, 4..[55] Vohl, D., Fluke, C.J., Barnes, D.G., Hassan, A.H., 2017. Real-time colouring and filtering with graphics shaders. MNRAS 471,3323–3346.[56] Wen, Z., O’dell, C.R., 1995. A three-dimensional model of theOrion Nebula. ApJ 438, 784–793.[57] Westover, L.A., 1991. Splatting: A parallel, feed-forward vol-ume rendering algorithm. Ph.D. thesis. Chapel Hill, NC, USA.UMI Order No. GAX92-08005.[58] York, D.G., Adelman, J., Anderson, John E., J., Anderson,S.F., Annis, J., Bahcall, N.A., Bakken, J.A., Barkhouser, R.,Bastian, S., Berman, E., Boroski, W.N., Bracker, S., Briegel, C.,Briggs, J.W., Brinkmann, J., Brunner, R., Burles, S., Carey, L.,Carr, M.A., Castander, F.J., Chen, B., Colestock, P.L., Con-nolly, A.J., Crocker, J.H., Csabai, I., Czarapata, P.C., Davis,J.E., Doi, M., Dombeck, T., Eisenstein, D., Ellman, N., Elms,B.R., Evans, M.L., Fan, X., Federwitz, G.R., Fiscelli, L., Fried-man, S., Frieman, J.A., Fukugita, M., Gillespie, B., Gunn, J.E.,Gurbani, V.K., de Haas, E., Haldeman, M., Harris, F.H., Hayes,J., Heckman, T.M., Hennessy, G.S., Hindsley, R.B., Holm, S.,Holmgren, D.J., Huang, C.h., Hull, C., Husby, D., Ichikawa,S.I., Ichikawa, T., Ivezi´c, ˇZ., Kent, S., Kim, R.S.J., Kinney,E., Klaene, M., Kleinman, A.N., Kleinman, S., Knapp, G.R.,Korienek, J., Kron, R.G., Kunszt, P.Z., Lamb, D.Q., Lee, B.,Leger, R.F., Limmongkol, S., Lindenmeyer, C., Long, D.C.,Loomis, C., Loveday, J., Lucinio, R., Lupton, R.H., MacKin-non, B., Mannery, E.J., Mantsch, P.M., Margon, B., McGehee,P., McKay, T.A., Meiksin, A., Merelli, A., Monet, D.G., Munn,J.A., Narayanan, V.K., Nash, T., Neilsen, E., Neswold, R., New-berg, H.J., Nichol, R.C., Nicinski, T., Nonino, M., Okada, N.,Okamura, S., Ostriker, J.P., Owen, R., Pauls, A.G., Peoples,J., Peterson, R.L., Petravick, D., Pier, J.R., Pope, A., Pordes,R., Prosapio, A., Rechenmacher, R., Quinn, T.R., Richards,G.T., Richmond, M.W., Rivetta, C.H., Rockosi, C.M., Ruth-mansdorfer, K., Sand ford, D., Schlegel, D.J., Schneider, D.P.,Sekiguchi, M., Sergey, G., Shimasaku, K., Siegmund, W.A.,Smee, S., Smith, J.A., Snedden, S., Stone, R., Stoughton, C.,Strauss, M.A., Stubbs, C., SubbaRao, M., Szalay, A.S., Sza-pudi, I., Szokoly, G.P., Thakar, A.R., Tremonti, C., Tucker,D.L., Uomoto, A., Vanden Berk, D., Vogeley, M.S., Waddell,P., Wang, S.i., Watanabe, M., Weinberg, D.H., Yanny, B., Ya-suda, N., SDSS Collaboration, 2000. The Sloan Digital SkySurvey: Technical Summary. AJ 120, 1579–1587.[59] ZuHone, J.A., Kowalik, K., Ahman, E., Lau, E., Nagai, D.,2018. The galaxy cluster merger catalog: An online repositoryof mock observations from simulated galaxy cluster mergers.ApJS 234, 4.