How well does magphys recover galaxy properties? A test using eagle simulated star-forming galaxies 2026 How well does magphys recover galaxy properties? A test using eagle simulated star-forming galaxies – A Zoe R. Jones, Elisabete da Cunha, Andrew Battisti Abstract Spectral energy distribution (SED) models are widely used to infer the physical properties of galaxies from multi-wavelength photometry, but their accuracy is difficult to assess because the true properties of observed galaxies are generally unknown. We address this by fitting synthetic SEDs of ∼ 31,000 \sim 31,000 star-forming galaxies drawn from the eagle cosmological simulations, post-processed with the skirt radiative transfer code, using the magphys SED modelling framework. This provides a controlled testbed with known intrinsic parameters, enabling a direct assessment of model accuracy and the origin of systematic biases. Under idealised conditions, fitting well-sampled ultraviolet-to-submillimetre SEDs at z = 0.1 z=0.1 , z = 2 z=2 , and z = 5 z=5 , magphys recovers stellar mass, star formation rate, specific star formation rate, dust mass, and dust luminosity to within ≲ 0.14 \lesssim 0.14 dex, while mass-weighted stellar ages are not robustly constrained. We find that mismatches between the assumed star formation history (SFH) priors and the intrinsic SFHs of the simulated galaxies introduce systematic biases in stellar mass estimates, even when the fits provide good statistical agreement. To assess performance under realistic survey conditions, we construct a WAVES-like mock sample using optical and near-infrared photometry with realistic uncertainties. In this case, stellar masses and star formation rates remain well constrained (systematic offsets ≲ 0.1 \lesssim 0.1 dex; scatters ≃ 0.07 \simeq 0.07 and ≃ 0.15 \simeq 0.15 dex, respectively), whereas dust properties degrade significantly without far-infrared data: dust luminosities show offsets of ≃ 0.30 \simeq 0.30 dex and scatters ≃ 0.25 \simeq 0.25 dex, and dust masses exhibit scatters ≃ 0.3 \simeq 0.3 dex. We conclude that MAGPHYS is a reliable tool for recovering key galaxy properties from broad-band photometry, but that SFH assumptions and limited wavelength coverage introduce significant uncertainties, particularly for dust and stellar ages. 1 Introduction The multi-wavelength emission, or spectral energy distributions (SEDs), of galaxies encode much of their critical information, such as their stellar masses and star formation rates (see reviews by, e.g., 102 ; 28 ; 70 ). However, interpreting these parameters from observed galaxy emission alone is complex, due to numerous degeneracies in SED colours that arise from the complex combinations of emitting and attenuating sources (e.g., stellar age and dust attenuation in ultraviolet (UV) to optical colours). Nevertheless, various SED models have been developed over the past decade and a half that aim to robustly model the observed SEDs of galaxies at various redshifts, and constrain their physical properties (e.g., 35 ; 79 ; 25 ; 20 ; 60 ; 88 ; see recent review by 81 ). The vast majority of these SED models include a spectral synthesis model for the emission of stellar populations (e.g., 12 ), combined with some assumption on the stellar initial mass function (IMF) and star formation history (SFH), a dust attenuation prescription (e.g., 15 ; 14 ; 23 ; see 90 for a review), and an energy balance method to connect the UV, optical, and near-infrared stellar emission consistently to the mid- and far-infrared dust emission (e.g., 35 ). SED models therefore rely heavily on simplifications and assumptions to model galaxy emission, with the goal of restricting the parameter space to save computing time while still being realistically applicable to large data samples. The exact computation of true-to-life galaxy SEDs would require more computationally-intensive radiative transfer models (e.g., 3 ; 86 ; 75 ), which are at present not easily applicable to the extraction of physical parameters from large galaxy surveys. A main challenge of frequently used SED models comes from determining how reliable their recovered parameters are. Several studies have compared some of the available SED models with each other to check for systematic differences between different codes (e.g., 59 ; 81 ), while others have compared the more fundamental stellar population models, dust models, and photometry choices, and their effects on parameter recovery (e.g., 71 ; 96 ; 99 ). However, this still leaves us with the question of how well SED models recover the true physical parameters of observed galaxies. That is, such studies sought to find which SED codes agree most with each other in order to predict their accuracies, rather than comparing them to a ground truth. Given the degeneracies in galaxy SEDs, and the relatively large number of free parameters in these models compared to the number of observables (usually, multi-wavelength broad-band fluxes), simply checking the goodness-of-fit to the observations is not sufficient to tell us whether a model fit is robust or not, and whether we can trust the recovered parameters. SED models try to overcome these degeneracies and account for uncertainties by adopting Bayesian frameworks that produce the full likelihood distributions of the physical parameters, and allow one to assess the impact of degeneracies and uncertainties. Still, the absolute calibration of the recovered parameters remains a challenge. One way to test the outputs of SED models is to apply them to galaxies for which we know the ‘ground truth’ using other methods. For example, 9 developed the photometric redshift extension of the magphys model ( 35 ; 34 ) and, when testing this code on galaxies with known spectroscopic redshifts from the COSMOS survey ( 95 ; 104 ) , they found a better agreement between the magphys photometric redshifts and the spectroscopic redshifts when a 2175 Å feature was included in the model dust attenuation curves, rather than using the featureless power-law model of 23 . This proved to be an independent way of testing SED model assumptions on real galaxies. Another way to do this is to use mock galaxies, as in the work of 57 , who used the sunrise Monte Carlo dust radiative transfer code ( 63 ; 62 ) to create synthetic SEDs of two gadget-3 ( 97 ; 98 ) simulated galaxies for which the physical parameters are known: an isolated disk galaxy, and a merger of two disk galaxies. They then used the magphys model to fit these synthetic SEDs and constrain the parameters of the simulated galaxies, and compared them to the ‘true’ known parameters (see also 46 for a similar test in the context of recovering the physical parameters of sub-millimetre galaxies). 56 conducted a similar study where they fit the synthetic SEDs of the nihao+skirt catalogue ( 103 ; 49 ) with the provabgs SED code ( 55 ; 64 ) , finding the impact of assumed dust geometry on parameter recovery. 74 conducted a similar study at high-redshift using galaxies from the simba simulations ( 37 ) , processed by the powerday dust radiative transfer package ( 75 ) , and fitted by the prospector SED model ( 61 ; 60 ) . They found that, at z\sim 7 , outshining by young stellar populations significantly affects recovered stellar masses. Using synthetic SEDs from simulations provides a ‘ground truth’ to which the SED-recovered physical parameters can be directly compared. In this way, the accuracy of the SED code is determined not in relation to other codes but to the ‘ground truth’ itself. That is the approach taken in this paper. Here, we investigate whether the magphys model can accurately recover the physical parameters of simulated galaxies from their synthetic SEDs by using a large sample of galaxies from the eagle simulations, for which UV-to-IR SEDs were computed by 17 ; 18 using the skirt radiative transfer code. The eagle simulations provide us with a much larger sample of galaxies (compared to the study of 57 ) that cover a wider parameter space, which allows us to test magphys more widely and investigate more thoroughly any potential systematic uncertainties and, most importantly, what causes them. For each galaxy, we have an ‘observed’ SED from skirt that we can then fit with magphys , and we know the ‘true’ physical parameters of these galaxies, that we can compare with the ones recovered by magphys . Thus, we enhance the statistical power of previous studies by using a sample of 30,928 simulated galaxies in three redshift snapshots, z=0.10 , z=2.01 , and z=5.04 . We begin by testing magphys under idealised conditions, fitting the full ultraviolet-to-submillimetre SEDs computed by skirt in fifty photometric bands. We then assess the performance of magphys in more realistic observational scenarios by constructing a mock galaxy sample with photometric coverage and uncertainties matched to those anticipated for the upcoming Wide-Area VISTA Extragalactic Survey (WAVES; 45 ), thereby evaluating the reliability of parameter recovery in a survey-like context. We note that any such tests are necessarily limited by the fidelity of the simulations used as ‘ground truth’, in particular by the assumptions and resolution limitations inherent to radiative transfer modelling. Our aim is not to claim that eagle + skirt provides a perfect representation of real galaxies, but rather to use it as a controlled testbed in which the intrinsic physical parameters are known. This enables us to isolate the impact of SED fitting assumptions on parameter recovery in a way that is not possible using observational data alone. In this work, we restrict our analysis to star-forming galaxies and do not consider systems dominated by active galactic nuclei (AGN). This paper is organised as follows. In Sect. 2 , we describe the eagle simulations, the skirt radiative transfer code, the simulated galaxy samples and the magphys model. In Sect. 3 , we present the comparisons between the true physical parameters of the simulated galaxies and the parameters recovered by magphys . In Sect. 4 we discuss these results as well as the systematic uncertainties we discover, their potential sources, and ways to reduce them. We summarise and conclude our paper in Sect. 5 . Additional results are presented in Appendix A . Throughout this paper, we use AB magnitudes ( 80 ) , and assume the standard cosmological parameters from Planck ( 83 ; 84 ) , with the key parameters of \Omega_{m}=0.307 , \Omega_{\Lambda}=0.693 , \Omega_{b}=0.04825 , h=0.6777 , and \sigma_{8}=0.8288 . 2 Methods 2.1 The Mock Galaxy Sample 2.1.1 The eagle + skirt Simulated Galaxies We use a sample of mock galaxies from the Evolution and Assembly of GaLaxies and their Environments ( eagle ) simulations that were post-processed by 17 using the Stellar Kinematics Including Radiative Transfer ( skirt ) radiative transfer code, and released as a database of UV to submillimetre broadband fluxes by 18 . eagle ( 30 ; 93 ) is a suite of hydrodynamical simulations that follow the formation of large-scale structure and galaxies in a \Lambda CDM Universe. The simulations were run using a modified version of the gadget-3 smooth particle hydrodynamical (SPH) code ( 98 ) . This modified hydrodynamics scheme, known as ANARCHY ( 93 ; 92 ) , affects the implementation of SPH, time-stepping, and subgrid models of the simulations that govern cooling, star formation, stellar evolution, feedback from star formation, gas accretion on to and mergers of black holes, and active galactic nucleus (AGN) feedback. Within these simulations, galaxies are defined as gravitationally-bound sub-halos identified through use of the friends-of-friends (FoF) algorithm ( 38 ) and the SUBFIND algorithm ( 97 ; 43 ) . Galaxy properties are then calculated from the relevant particles in the corresponding sub-halo and that are within a 3D aperture with a radius of 30 proper kpc (pkpc) from the galaxy’s stellar centre of mass. The subgrid physics used in the eagle simulations is based on the physics implemented in the OverWhelmingly Large (OWL) Simulations ( 94 ) , and is described in detail in 93 . The eagle simulations provide a realistic set of simulated galaxies, as they have been shown to reproduce the stellar mass functions and optical colours of local galaxies ( 51 ; 100 ) . Following the description of the eagle simulations in 93 and 30 , 72 presents a database of these simulations for public use. skirt is a 3D Monte Carlo radiative transfer code that simulates the transfer of radiation through dusty media ( 4 ; 7 ; 16 ) . It follows the lives of photons as they travel through the interstellar medium (ISM), where they are absorbed or scattered by dust, and computes the heating of and re-emission by dust grains self-consistently. It can be used to consistently model the total stellar and dust emission of galaxies (e.g., 3 ; 40 ). Later updates to skirt allow the code to self-consistently calculate the dust emission of galaxies, assuming the dust content of said galaxies are in local thermal equilibrium (LTE; 5 ; 6 ). Early versions of the skirt code used the 108 dust model only to describe the dust optical properties of galaxies ( 3 ) , however later updates have added multiple different dust models that can be chosen to use, including the 44 dust model which is built into the DustME library ( 27 ) . By linking the skirt code to DustME, a numerical tool designed to calculate the emission of dust assuming LTE, the skirt code also calculates the dust temperature distribution of a galaxy and the associated FIR and submm emission, including that from transiently heated dust grains and PAHs molecules ( 7 ) . This is possible even for galaxies with various custom mixtures of dust grains, dust distributions, and geometry of stellar components, which can all be selected and combined in as the user desires ( 16 ; 2 ) . 17 used skirt to post-process a sample of eagle galaxies with properties that match the selection criteria of local dusty galaxies from the Herschel Reference Survey ( 11 ; 29 ) . For the skirt processing, they selected the ZubkoDustMix dust mixture ( 108 ) and the OctTreeDustGrid spatial distribution of the dusty medium ( 89 ) . They overcame the resolution limitations of the simulations by employing sub-grid models for the star-forming regions and for the diffuse dust distribution in the eagle galaxies which are thoroughly detailed in 17 . skirt produces the UV to submm emission for each galaxy in different photometric bands to obtain fluxes that can be directly compared with observations of galaxies using various telescopes and instruments. 17 then used these mock observations to show that eagle can reproduce the infrared and submm properties of local galaxies. 18 later extended this work to a larger sample of nearly half a million simulated eagle galaxies with \log(\hbox{$M_{\ast}$}/\hbox{$M_{\odot}$})>8.5 and sufficiently resolved dust (with the number of numerical particles representing the dust content in a galaxy N_{\mathrm{dust}}>250 ) in 23 redshift snapshots up to redshift z=6 , across six eagle runs. They found 316,389 galaxies that produced sufficiently resolved dust distributions to calculate meaningful dust-attenuated stellar emission and dust emission fluxes. The galaxies retain their original viewing angles from eagle , which are labelled as ‘random’ viewing angles, as 18 also considers edge-on and face-on viewing angles at UV and optical wavelengths. They then extended the eagle database by publishing the rest-frame magnitudes and observer-frame fluxes of these 316,389 galaxies in 50 standard photometric filters from UV to mm wavelengths ( \sim 0.15 to \sim 1250 \mu m). These bands are: GALEX FUV/NUV ( 73 ) ; SDSS ugriz ( 42 ) ; 2MASS JHKs ( 26 ) ; UKIDSS ZYJHK ( 65 ) ; Johnson UBVRIJM ( 10 ) ; WISE W1/W2/W3/W4 ( 106 ) ; IRAS 12/25/60/100 \mu m ( 78 ) ; Spitzer /IRAC I1/I2/I3/I4 ( 50 ) ; Spitzer /MIPS 24/70/160 \mu m ( 87 ) ; Herschel /PACS 70/100/160 \mu m ( 85 ) ; Herschel /SPIRE 250/350/500 \mu m ( 54 ) ; SCUBA2 450/850 \mu m ( 58 ) ; and ALMA Bands 10/9/8/7/6 ( 1 ) . The names and pivot wavelengths of these photometric filters are displayed in Table 1 , and ordered from shortest to longest wavelength. These magnitudes and fluxes are given in the ‘DustyMagnitudes’ and ‘DustyFluxes’ tables of the database, respectively. The ‘DustFit’ table was also added to the database to provide the computed dust temperature and dust mass of each galaxy. Filter name \lambda pivot ( \mu m) GALEX_FUV 0.1535 GALEX_NUV 0.2301 Johnson_U 0.3525 SDSS_u 0.3557 Johnson_B 0.4417 SDSS_g 0.4702 Johnson_V 0.5525 SDSS_r 0.6176 Johnson_R 0.6899 SDSS_i 0.7490 Johnson_I 0.8739 UKIDSS_Z 0.8826 SDSS_z 0.8947 UKIDSS_Y 1.031 TwoMASS_J 1.239 Johnson_J 1.243 UKIDSS_J 1.250 UKIDSS_H 1.649 TwoMASS_H 1.649 TwoMASS_Ks 2.164 UKIDSS_K 2.206 Filter name \lambda pivot ( \mu m) WISE_W1 3.390 IRAC_I1 3.551 IRAC_I2 4.496 WISE_W2 4.641 Johnson_M 5.012 IRAC_I3 5.724 IRAC_I4 7.884 IRAS_12 11.41 WISE_W3 12.57 WISE_W4 22.31 IRAS_25 23.61 MIPS_24 23.76 IRAS_60 60.41 PACS_70 70.77 MIPS_70 71.99 PACS_100 100.8 IRAS_100 101.1 MIPS_160 156.4 PACS_160 161.9 SPIRE_250 252.5 ALMA_10 349.9 SPIRE_350 354.3 SCUBA2_450 449.3 ALMA_9 456.2 SPIRE_500 515.4 ALMA_8 689.6 SCUBA2_850 853.8 ALMA_7 937.9 ALMA_6 1244 Table 1: The names and pivot wavelengths of the 50 parametric bands for which mock broadband observed fluxes and absolute AB magnitudes are available in the eagle database. The left table list the UV and optical filters, for which face-on, edge-on and random orientation measurements are available. The right table lists the IR and submm filters which only have one orientation available, as emission in these bands is essentially isotropic. 2.1.2 The idealised subsample of mock galaxies To build our idealised mock galaxy sample, we use star-forming galaxies from three redshift snapshots provided by the eagle database: snapshots 27 ( z=0.10 ), 15 ( z=2.01 ) and 8 ( z=5.04 ); we chose these snapshots as representative of the low-, intermediate-, and high-redshift Universe, respectively. We selected galaxies from five of the six available eagle models in the database: Ref-L0025N0752, Recal-L0025N0752, Ref-L0025N0376, Ref-L0050N0752, and Ref-L0100N1504. These models differ in: subgrid physics assumptions, indicated by the ‘Ref’ or ‘Recal’ prefix; box-length of the simulation, with each box-length given in comoving Mpc by the four digits following ‘L’; and total number of both the baryonic and dark matter particles, the cubed root of which is given by the four digits following ‘N’. We exclude galaxies from the sixth eagle model, AGNdT9-L0050N0752, as we are only interested in testing the magphys model without AGN templates for now. This allows us to create three samples of mock galaxies at three redshifts. The distribution of these galaxies by model and redshift is shown in Table 2 . eagle Model Galaxies at Galaxies at Galaxies at z=0.10 z=2.01 z=5.04 Ref-L0025N0752 527 298 28 Recal-L0025N0752 384 233 26 Ref-L0025N0376 155 227 18 Ref-L0050N0752 1150 2003 223 Ref-L0100N1504 8072 15445 2139 Combined Total 10288 18206 2434 Table 2: The eagle galaxies that are included in our three samples used when testing the magphys model, displayed by eagle model and redshift. In total, we investigated 30,928 galaxies. We obtained the SEDs and parameters of these galaxies from the eagle database. Specifically from 18 we obtained: • the observer-frame, dust-affected fluxes at 50 wavelength filters ranging from \sim 0.15 to \sim 1250 \mu m from the ‘DustyFluxes’ table; • and the representative dust temperatures and dust masses, provided as ‘Temp_Dust’ and ‘Mass_Dust’ on the ‘DustFit’ table. From ( 72 ) we obtained: • the average, birth mass-weighted stellar population ages, provided as ‘InitialMassWeightedStellarAge’ on the ‘SubHalo’ table; • the total current masses (baryonic and dark matter), the gas masses, and stellar masses, provided as ‘Mass’, ‘MassType_Gas’ and ‘MassType_Star’ on the ‘SubHalo’ table; • the total star formation rates, provided as ‘StarFormationRate’ on the ‘SubHalo’ table; • the stellar metallicities, provided as ‘Stars_Metallicity’ on the ‘SubHalo’ table; • and the stellar half-mass radii within an aperture of 30 pkpc, provided as ‘R_halfmass30’ on the ‘Sizes’ table. Figure 1: Normalised distributions of nine parameters of the 10,288 galaxies at redshift z=0.10 (shown in blue), 18,206 galaxies at redshift z=2.01 (shown in green), and 2,434 galaxies at redshift z=5.04 (shown in red) of our idealised mock galaxy sample, and 35,051 galaxies at redshift z\leq 0.27 (shown in black) of our WAVES-like sample. The nine parameters displayed are: (a) stellar mass, (b) dust mass, (c) dust luminosity, (d) SFR, (e) sSFR, (f) average stellar age, (g) stellar metallicity, (h) dust temperature, and (i) half-mass radius. The vertical lines show the median of each distribution. Figure 2: Median synthetic SEDs of the three redshift samples (normalised in the K -band), showing the 16 th and 84 th percentile values as dotted lines, and the 50 th percentile values as solid lines. The redshift z=0.10 galaxies are shown in blue, the redshift z=2.01 galaxies are shown in green, and the redshift z=5.04 galaxies are shown in red. Fig. 1 displays the normalised distributions of nine key parameters of the galaxies in our mock sample: stellar mass, dust mass, dust luminosity, SFR, sSFR, average stellar age, stellar metallicity, dust temperature and half-mass radius. Our local ( z=0.1 ) mock galaxy sample includes higher stellar mass galaxies with higher metallicities, older ages, and lower specific star formation rates, as expected from a more evolved population of galaxies in the local Universe. The high-redshift ( z=5.04 ) sample extends to lower masses and metallicities, younger ages, and higher specific star formation rates. We note the very narrow range of average stellar ages in each of the mock galaxy subsamples (panel f): the subsamples do not overlap in ages at all. While the stellar ages in galaxies at higher redshifts are limited by the age of the Universe at those redshifts, younger average ages are possible at low redshifts if there is significant recent star formation; we will return to this issue in Section 4 . We also display the K -band normalised median SEDs for each of the three subsamples in Fig. 2 , which shows the average shape of the synthetic SEDs. We note the relative lack of variability of the synthetic SEDs in the z=0.1 sample compared to the higher redshift subsamples. On average, the local SEDs display redder UV/optical colours, as a result of their relatively older ages combined with dust attenuation. Their infrared SEDs peak at longer wavelengths as a result of the colder dust temperatures, as shown in Fig. 1 (h). It is interesting to note that the mock galaxies seem to display much larger variation of their optical-to-IR SEDs in the highest-redshift subsample. 2.1.3 The WAVES-like subsample of mock galaxies To evaluate the performance of magphys under realistic observational conditions, we construct an additional subsample of mock galaxies designed to mimic the properties of the upcoming Wide-Area VISTA Extragalactic Survey (WAVES; 45 ). WAVES will obtain spectroscopy for approximately 1.6 million galaxies using the 4-metre Multi-Object Spectroscopic Telescope (4MOST; 39 ), divided into two components: WAVES Wide and WAVES Deep. In this work, we focus on the WAVES Wide survey, which aims to obtain spectra for approximately 0.9 million galaxies over an area of \sim 1200 square degrees. For the majority of WAVES Wide galaxies, broad-band ugriZYJHK photometry will be available from the KiDS ( 105 ) and VIKING ( 48 ) imaging surveys. The survey primarily targets galaxies at redshifts z\lesssim 0.2 , with a limiting Z -band magnitude of Z\leq 21.1 . To reproduce the redshift range probed by WAVES Wide, we extend our idealised low-redshift sample by including additional eagle snapshots at z=0.00 , 0.18 , and 0.27 , drawn from the same five eagle simulation runs used previously. Including galaxies up to z=0.27 allows us to capture the effective redshift range expected for WAVES Wide once photometric redshift uncertainties are taken into account. This increases the size of the low-redshift mock sample from 10,288 galaxies at z=0.10 to 42,588 galaxies at z\leq 0.27 . We then apply a Z -band magnitude cut of Z\leq 21.1 , matching the WAVES Wide selection criteria, which yields a final WAVES-like sample of 35,051 galaxies. The physical properties of these mock galaxies are plotted in Fig. 1 . For this sample, we fit only the nine ugriZYJHK photometric bands available to WAVES, rather than the full set of 50 bands provided in the eagle database. To assign realistic photometric uncertainties, we adopt the limiting magnitudes reported in Table 1 of 105 and convert them into flux uncertainties assuming Gaussian errors. This WAVES-like subsample therefore enables us to assess the accuracy and robustness of magphys parameter recovery in a survey-like observational context. 2.2 SED fitting with magphys The magphys SED modelling code ( 35 ; 34 ) can interpret the UV, optical, and near-infrared emission from stellar populations in galaxies consistently with their dust emission in the infrared regime. magphys uses the stellar population synthesis models of 12 to compute light emitted by stellar populations of various ages and metallicities in galaxies between 91 Å and 160\mu m. The 22 IMF is assumed. The stellar emission from galaxies is computed by combining these stellar populations with a range of star formation histories that rise linearly at early ages and then decline exponentially, with varying timescales and superimposed random bursts of star formation to account for stochasticity in the star formation histories. This starlight is then attenuated by dust using the two-component model of 23 , which accounts for newly-formed stars ( <10 Myr) inside their birth clouds being more attenuated than older stars in the diffuse ISM. The effective dust attenuation curves in the birth cloud and diffuse ISM components have different slopes: these are steeper in the birth cloud component ( \tau_{\lambda}\propto\lambda^{-1.3} ) than in the diffuse ISM ( \tau_{\lambda}\propto\lambda^{-0.7} ). This produces an age-dependent effective dust attenuation curve (see more details in 23 ; 35 ; 8 ). By combining the emission from the stellar populations with this dust attenuation model, magphys computes the total amount of energy absorbed by dust in both stellar birth clouds and the diffuse ISM and, through an energy balance argument, assumes that emission must be re-radiated by dust grains in the mid- to far-infrared. This emission is modelled using an empirical template for the polycyclic aromatic hydrocarbon (PAH) mid-infrared features, a hot dust component for the mid-infrared continuum, and two modified black bodies for the far-infrared continuum, which have dust equilibrium temperatures that are kept as free parameters. The relative contributions of these different dust emission components to the total dust emission are also free parameters. To fit the observed SEDs of galaxies, magphys generates a random library of stellar population models, for various star formation histories, metallicities and dust content, and a random library of dust emission models, for various dust temperatures and contributions by different dust components. These libraries are designed to encompass all plausible parameter combinations of the observed SED. The final stellar population library and the final infrared spectra library both consist of 50,000 different models each. The models in the two libraries are associated with each other through the parameter f_{\mu} , which is the fraction of the total infrared luminosity contributed by dust in the ISM. For each association with a sufficiently similar f_{\mu} value ( \Delta f_{\mu}\leq 0.15 ), the combined stellar and infrared spectra is included in the final model SED library. The final library of combined UV, optical, and IR SEDs consists of \sim 661 million models. This large number of models is required to properly sample the multi-dimension observational space. magphys compares the observed flux in each photometric band with the flux predicted by each model in the library and computes the \chi^{2} goodness-of-fit and probability of each model. Then, for each parameter, it marginalises the probability over all other parameters to produce a posterior likelihood distribution. We adopt the magphys estimated parameter value to be the median of that likelihood distribution, and the confidence range to be its 16th-84th percentile range. We formatted the flux measurements for each galaxy described in Sect. 2.1.2 when observed at random viewing orientations as magphys inputs. In total, we fit the UV-to-submm SEDs of 30,928 mock galaxies at redshifts z=0.10 , z=2.01 , and z=5.04 from the eagle Ref-L0025N0752, Recal-L0025N0752, Ref-L0025N0376, Ref-L0050N0752, and Ref-L0100N1504 simulations. We did not add noise to the photometric data, and instead applied a uniform uncertainty of 20% to each flux measurement, in order to investigate potential systematic uncertainties present within magphys rather than those caused by noisy data. For the redshift z=0.10 subsample, we use the magphys priors described in 35 , whereas for the redshift z=2.01 and z=5.04 subsamples, we use the magphys -highz version ( 34 ) , which includes priors that are more appropriate for high-redshift galaxies (specifically, a more diverse range of SFHs, and an extension to higher dust attenuations and warmer dust temperatures). For each mock galaxy, magphys fits the input SED and returns a best-fit model along with the likelihood distributions of key parameters. Here we focus on the following parameters: • stellar mass, M_{\ast} ; • star formation rate (averaged over the last 100 Myr), SFR; • specific star formation rate, {\rm sSFR}={\rm SFR}/\hbox{$M_{\ast}$} ; • total dust luminosity, L_{\rm dust} ; • total dust mass, M_{\rm dust} ; • mass-weighted age, age M . We compare these parameters directly with the ‘true’ values provided by eagle + skirt in Sect. 3 . In some cases, the eagle + skirt database does not provide the parameter of interest, or the magphys and eagle + skirt parameters are defined in slightly different ways. We discuss these cases in detail when presenting their results. 2.3 Comparison metrics In our analysis, we quantify the accuracy of the recovered magphys parameter values by measuring the median difference or offset between them and the true eagle values, \Delta , given as: \Delta=\text{median}(x_{\text{\hbox{{magphys}}}}-x_{\text{\hbox{{eagle}}}})\,, (1) where x_{\text{\hbox{{magphys}}}} and x_{\text{\hbox{{eagle}}}} correspond to the recovered logarithmic magphys values and true logarithmic eagle values, respectively. We additionally find for each parameter the standard deviations, lower quartiles (Q1), upper quartiles (Q3), median error bars, and normalised median absolute deviation (NMAD) of each parameters, with NMAD defined as: \text{NMAD}=1.4826\times\text{median}(|(x_{\text{\hbox{{magphys}}}}-x_{\text{\hbox{{eagle}}}})-\Delta|)\,, (2) where NMAD provides a robust estimate of the scatter around the median trend, approximately equivalent to the standard deviation for a Gaussian distribution but less sensitive to outliers. Finally, we find the outlier fraction of each parameter, where an outlier value is either lower than Q1-1.5\times IQR or higher than Q3+1.5\times IQR , and IQR is the interquartile range. 3 Results Parameter eagle Database Name Median Offset Standard Deviation NMAD Offset Q1 Offset Q3 Outlier Fraction Median Error Bar Redshift z=0.10 Subsample \log(\hbox{$M_{\ast}$}/\hbox{$M_{\odot}$}) MassType_Star -0.140 0.078 0.066 -0.190 -0.099 0.031 0.210 \log(\rm{SFR}/\hbox{$M_{\odot}$}\,\rm{yr}^{-1}) StarFormationRate -0.064 0.110 0.094 -0.132 -0.005 0.025 0.115 \log(\rm{sSFR}/\rm{yr}^{-1}) - 0.072 0.138 0.134 -0.016 0.165 0.012 0.270 \log(\hbox{age${}_{M}$}/\rm{yr}) InitialMassWeightedStellarAge -0.256 0.143 0.121 -0.356 -0.183 0.029 0.345 \log(\hbox{$L_{\rm dust}$}/L_{\odot}) - 0.014 0.038 0.035 -0.010 0.037 0.024 0.079 \log(\hbox{$M_{\rm dust}$}/\hbox{$M_{\odot}$}) Mass_Dust 0.071 0.058 0.049 0.036 0.103 0.020 0.090 Redshift z=2.01 Subsample \log(\hbox{$M_{\ast}$}/\hbox{$M_{\odot}$}) MassType_Star -0.050 0.142 0.066 -0.101 -0.010 0.106 0.075 \log(\rm{SFR}/\hbox{$M_{\odot}$}\,\rm{yr}^{-1}) StarFormationRate -0.027 0.098 0.058 -0.065 0.013 0.047 0.045 \log(\rm{sSFR}/\rm{yr}^{-1}) - 0.022 0.165 0.109 -0.045 0.105 0.067 0.110 \log(\hbox{age${}_{M}$}/\rm{yr}) InitialMassWeightedStellarAge -0.096 0.172 0.135 -0.193 -0.010 0.041 0.120 \log(\hbox{$L_{\rm dust}$}/L_{\odot}) - -0.066 0.033 0.016 -0.078 -0.055 0.062 0.025 \log(\hbox{$M_{\rm dust}$}/\hbox{$M_{\odot}$}) Mass_Dust 0.002 0.071 0.059 -0.041 0.040 0.041 0.160 Redshift z=5.04 Subsample \log(\hbox{$M_{\ast}$}/\hbox{$M_{\odot}$}) MassType_Star -0.070 0.203 0.114 -0.163 -0.002 0.109 0.035 \log(\rm{SFR}/\hbox{$M_{\odot}$}\,\rm{yr}^{-1}) StarFormationRate -0.019 0.145 0.073 -0.075 0.026 0.081 0.020 \log(\rm{sSFR}/\rm{yr}^{-1}) - 0.059 0.210 0.148 -0.034 0.171 0.045 0.030 \log(\hbox{age${}_{M}$}/\rm{yr}) InitialMassWeightedStellarAge -0.112 0.223 0.148 -0.259 -0.031 0.073 0.020 \log(\hbox{$L_{\rm dust}$}/L_{\odot}) - -0.056 0.041 0.026 -0.074 -0.039 0.044 0.020 \log(\hbox{$M_{\rm dust}$}/\hbox{$M_{\odot}$}) Mass_Dust -0.002 0.078 0.047 -0.031 0.034 0.085 0.270 Table 3: The six parameters we investigated to determine the accuracy of the magphys model. The columns show parameter names and units, parameter names used in the eagle database (if they are provided), median offsets, standard deviations, NMAD, Q1 offsets, Q3 offsets, outlier fractions, and median error bars of the recovered magphys values when compared to the true eagle values. Our sample has been separated by redshift subsamples for ease of comparison. 3.1 Quality of the fits Figure 3: The SEDs of three of the eagle galaxies when fit by magphys . The observed fluxes at each wavelength are shown as red dots, the unattenuated stellar emission are shown in blue, and the SEDs are shown in black, with the predicted fluxes at each wavelength shown as open, black circles. The galaxy ID is shown in the top left corner of each panel. The redshift of the galaxy and the \chi^{2} value of each fit is shown in the top right corner of each panel. The residuals for each fit are shown below the SEDs, and share the same definition as the residuals shown in Fig 4 . The wavelengths shown are in the observer frame of reference. We first check how well magphys fits the synthetic SEDs provided by eagle . Fig. 3 presents the SEDs of three of the eagle galaxies at redshift z=0.10 . The eagle galaxy with ID 20574262, shown in panel (a), was the best-fit galaxy of our redshift z=0.10 subsample, with the lowest \chi^{2} . Most of the observed fluxes for this galaxy lie exactly on the curve predicted by magphys , which indicates the high quality of this fit. Panel (b) displays the fit for eagle galaxy 16382118, which has a fit \chi^{2} value close to the median of the whole low- z subsample. This is still a very good fit, as shown by the small residuals across wavelength. Panel (c) displays the fit for eagle galaxy 11581940, one of the worst-fit galaxies. Even with a relatively large \chi^{2} , most of the flux observations of this galaxy are matched very closely to the model SED. The fluxes that are not matched well are primarily in the FIR, which could be attributed to the differences in the dust temperatures of the models: Fig. 1 shows that some eagle galaxies have dust temperatures below 15 K, which is the coldest dust temperature allowed by the magphys prior. Figure 4: The median SED residuals for the three redshift samples, with redshift z=0.10 galaxies shown in panel a) in blue, redshift z=2.10 galaxies shown in panel b) in green, and redshift z=5.04 galaxies shown in panel c) in red. The error bars of all three panels show the 16 th and 84 th percentiles of the residuals, which share the same definition as the residuals shown in Fig 3 . The wavelengths shown are in the observer frame of reference. To investigate how well the SEDs are fitted in general, we plot in Fig. 4 , the median fit residuals in each band for the three redshift subsamples. The median differences between the observed and model flux values are very low across all wavelengths. The largest residuals are found in the observer-frame MIR regime, which could be due to the relatively simple PAHs emission template used by magphys in that regime. There are also significant residuals in the longest wavelength band at z=0.10 , which are likely due to the fact that some eagle galaxies extend to colder dust temperatures than allowed by the magphys prior. When assessing the reduced \chi^{2} values of our sample, which compares the flux measurements of the observed spectra to the model spectra at each photometric band, we find that 99.7% of the redshift z=0.10 galaxies, 98.3% of the redshift z=2.01 galaxies, and 73.5% of the redshift z=5.04 galaxies have a \chi^{2} value of \leq 1 . The good agreement between the observed and model spectra indicates that the recovered parameters of magphys are well constrained; we now investigate their accuracy by comparing them to the known values. A summary of the parameters we investigated and the values of their recovered offsets with the true values are presented in Table 3 . These offsets are depicted in Fig. 5 and Fig. 6 , and discussed in more detail in the following sections. Figure 5: Comparisons between the true parameters of the eagle mock galaxies ( y -axes), and the median-likelihood parameters recovered by magphys from fitting their skirt -processed SEDs ( x -axes). The left-hand, middle, and right-hand columns show the comparison for the redshift z=0.10 , z=2.01 , and z=5.04 subsamples, respectively, for each parameter. The top row shows the comparisons of stellar masses, the middle row shows SFR, and the bottom row shows sSFR. All three parameters are coloured by the recovered magphys mass-weighted average stellar ages. The diagonal black line represents the one-to-one relation. Printed in the bottom right corners of the panels are the median offsets between the trends and the one-to-one relation, and the standard deviations of the trends. Printed in the top left corners of the panels are the median error bars of the recovered values, defined as the median difference between the 16 th and 84 th percentiles of each recovered value. Figure 6: Comparisons between the true parameters of the eagle mock galaxies ( y -axes), and the median-likelihood parameters recovered by magphys from fitting their skirt -processed SEDs ( x -axes). The left-hand, middle, and right-hand columns show the comparison for the redshift z=0.10 , z=2.01 , and z=5.04 subsamples, respectively, for each parameter. The top row shows the comparisons of average stellar ages, the middle row shows dust luminosities, and the bottom row shows dust masses. Average stellar age is coloured by the recovered magphys stellar masses, and dust luminosity and mass are coloured by the recovered magphys mass-weighted average stellar ages. The diagonal black line represents the one-to-one relation. Printed in the bottom right corners of the panels are the median offsets between the trends and the one-to-one relation, and the standard deviations of the trends. Printed in the top left corners of the panels are the median error bars of the recovered values, defined as the median difference between the 16 th and 84 th percentiles of each recovered value. 3.2 Idealised eagle sample 3.2.1 Stellar Masses We compare the magphys likelihood estimates of stellar mass with the true values for all eagle galaxies at z=0.10 , z=2.01 , and z=5.04 in panels (a), (b) and (c) of Fig. 5 , respectively, with each galaxy coloured by the magphys likelihood estimate of average stellar age. We note that there is a clear lower bound to the eagle stellar masses at \log(\hbox{$M_{\ast}$}/\hbox{$M_{\odot}$})=8.5 as 18 only produced the SEDs of galaxies with stellar masses larger than this value. While the estimated and true stellar masses are well correlated, panel (a) shows that there is a systematic underestimation of stellar mass of 0.140 dex on average by magphys for the redshift z=0.10 subsample, and that at the highest masses, this offset is increased. This systematic offset is within the median error bar for this subsample, however we investigate its potential causes in Sect. 4 . Panels (b) and (c) of Fig. 5 show that the median offset in recovered stellar masses decreases for the higher-redshift subsamples, reaching 0.050 and 0.070 dex for the z=2.01 and z=5.04 samples, respectively. However, the NMAD and standard deviation indicate similar or larger scatter, implying that while the bias decreases, the overall dispersion in the recovered masses does not decrease. We find that at these redshifts the galaxies with younger magphys estimated average stellar ages have higher offsets in stellar mass, while galaxies with older stellar ages are recovered much better. We discuss the impact of average stellar ages on the recovery of stellar masses in Sect. 4 . 3.2.2 Star Formation Rates Panels (d), (e) and (f) of Fig. 5 compare the star formation rates provided by the eagle database with the magphys likelihood estimates, again coloured by the magphys average stellar ages. We note that the eagle SFRs are instantaneous, and computed as the sums of the star formation rates of all gas particles within the galaxy at the time the galaxy is recorded; in contrast, the magphys SFRs are averaged over the last 100 Myr. We find that magphys recovers the SFRs of the galaxies well, with low median offset values for all three redshift subsamples, which is likely due to the fact that the mock SEDs are extremely well-sampled from the UV to the submillimetre, and that realistic SFHs are assumed by magphys ( 34 ; 81 ) . In all three redshift subsamples, there is a large degree of scatter evident in the plots. In the redshift z=2.01 and z=5.04 samples, the younger galaxies show more disagreement than older galaxies, though there is still considerable scatter found among the older galaxies, especially in the redshift z=0.10 and z=2.01 subsample. Overall, however, the median systematic offsets are low and smaller than their respective median error bars which indicates that magphys is recovering SFR well. 3.2.3 Specific Star Formation Rates The eagle database does not provide the specific star formation rates of galaxies, therefore we calculated the ‘true’ sSFR values by dividing the SFR of galaxies by their stellar masses. The comparisons between these values and the magphys likelihood estimates are presented in panels (g), (h) and (i) of Fig. 5 . All three redshift subsamples show a strong correlation between sSFR and the average stellar ages of galaxies, with high sSFR galaxies being younger and lower sSFR galaxies being older. The primary cause of the large scatter across the three redshift subsamples is likely inherited from the scatters of the stellar masses and SFRs, and the systematic uncertainties affecting these two parameters will most likely be the primary influence on the sSFR uncertainties as well. Even with these uncertainties included, the small median offset values lead us to conclude that magphys can recover sSFR to a satisfactory degree. 3.2.4 Stellar Ages The comparisons between the eagle values and the magphys likelihood estimates of mass-weighted stellar ages are presented in panels (a), (b), and (c) of Fig. 6 , and coloured by the magphys likelihood estimates of stellar mass. We find that there is a severe underestimation of a majority of the average stellar ages by magphys . For the redshift z=0.10 subsample, the median offset value of 0.261 dex and the median error bar value of 0.345 dex are the largest of the six parameters we investigated at any redshift. In the redshift z=2.01 and 5.04 subsamples, we find similar severe underestimations by magphys , despite the relatively low median offset values of 0.096 and 0.112 , respectively. In general, magphys can underestimate the stellar mass-weighted ages by up to a factor of 10 in the worst cases, though we also note the very small spread of the true values from the eagle galaxies. Due to the known difficulties of recovering the SFH of galaxies using only photometric methods (e.g., 21 ; 66 ), we investigated the cause of these underestimations and their impact on other parameters, specifically the recovered stellar masses, in Sect. 4 . 3.2.5 Dust Luminosities The eagle database does not provide the dust luminosities of its galaxies, therefore we calculated the ‘true’ value of L_{\rm dust} for each galaxy by directly integrating the SEDs computed by skirt between 8 and 1000 \mu m. We note that magphys defines L_{\rm dust} by integrating the dust emission SED between 3 and 1000 \mu m. We could not integrate the skirt SEDs in that exact range as they do not separate the dust and stellar emission, and the mid-IR continuum between 3 and 8 \mu m is severely contaminated by stellar emission. However, this should only introduce differences of at most a few percent in L_{\rm dust} . Panels (d), (e) and (f) of Fig. 6 show the comparisons between these ‘true’ values and the magphys likelihood estimates of dust luminosity, coloured by the magphys median-likelihood estimates of average stellar ages. The dust luminosity is the parameter derived with the most accuracy in our analysis. This is very likely thanks to the very well-sampled input SEDs, especially in the far-IR region where the dust emission peaks. In the redshift z=5.04 subsample, there is a visually obvious correlation between dust luminosity and mass-weighted stellar age, with the brightest dust luminosity galaxies being the youngest. Across all three redshift subsamples, the galaxies with older mass-weighted stellar ages have the highest scatter, however the overall scatter of all three redshift subsamples is very low. As the median offset values of all three redshift subsamples is lower than \sim 0.07 dex, we conclude that magphys recovers dust luminosities very well. 3.2.6 Dust Masses The comparisons between the eagle dust mass values and the magphys likelihood estimates of dust mass are presented in panels (g), (h) and (i) of Fig. 6 , and coloured by the magphys median-likelihood estimates of average stellar age. To compare these sets of values, we accounted for the difference in dust models used by 18 for the eagle database, and by the magphys model. 18 used the ( 108 ) dust model with an absorption coefficient at 350 \mu m of \kappa_{350}=0.330 cm 2 kg -1 to calculate the dust properties of the eagle galaxies, while magphys assumes the 47 value for an absorption coefficient at 850 \mu m of \kappa_{850}=0.077 cm 2 kg -1 . Therefore, we multiplied the recovered magphys dust mass values by a factor of 1.376 for all of the three redshift subsamples before comparing our results. Once that correction is made, we find that magphys recovers the dust masses of eagle galaxies very accurately. We note that 57 found a systematic underestimation of \sim 0.2-0.3 dex in their recovered magphys dust mass values. They attribute this to the assumptions of the cold dust phase of the ISM made by the simulations they investigated. We found no such underestimation in our results, with the largest median offset of our redshift subsamples being 0.071 dex, which occurs for the redshift z=0.10 subsample. In the redshift z=5.04 subsample, we find a strong correlation between dust mass and average stellar age, and find that the galaxies with younger stellar ages tend to have their higher dust masses overestimated slightly by magphys , however this trend is not present in the other redshift subsamples. Overall, the agreement is very good, and again it can be attributed to the good IR SED sampling of the mock galaxies. 3.3 WAVES-like sample Figure 7: Comparisons between the true parameters of the eagle mock galaxies ( y -axes), and the median-likelihood parameters recovered by magphys from fitting their skirt -processed SEDs for a WAVES-like filter set ( x -axes). The panels show the comparisons of a) stellar masses, b) average stellar ages, c) SFR, d) dust luminosities, e) sSFR, and f) dust masses. The parameters are coloured by recovered magphys average stellar age, except for average stellar age which is coloured by recovered magphys stellar mass. For ease of comparison, the grey dots show the sample of redshift z=0.10 galaxies fit using the adjusted age prior, as seen in Fig. 10 and Fig. 12 . The diagonal black lines represent the one-to-one relation. Printed in the bottom right corners of the panels are the median offsets between the trends and the one-to-one relation, and the standard deviations of the trends. Printed in the top left corners of the panels are the median error bars of the recovered values, defined as the median difference between the 16 th and 84 th percentiles of each recovered value. Parameter eagle Database Name Median Offset Standard Deviation NMAD Offset Q1 Offset Q3 Outlier Fraction Median Error Bar Adjusted Age Prior ( z=0.10 ) Subsample \log(\hbox{$M_{\ast}$}/\hbox{$M_{\odot}$}) MassType_Star -0.029 0.083 0.065 -0.073 0.015 0.042 0.140 \log(\rm{SFR}/\hbox{$M_{\odot}$}\,\rm{yr}^{-1}) StarFormationRate -0.071 0.112 0.086 -0.132 -0.015 0.040 0.105 \log(\rm{sSFR}/\rm{yr}^{-1}) - -0.047 0.134 0.115 -0.125 0.031 0.028 0.207 \log(\hbox{age${}_{M}$}/\rm{yr}) InitialMassWeightedStellarAge -0.081 0.073 0.068 -0.126 -0.034 0.018 0.225 \log(\hbox{$L_{\rm dust}$}/L_{\odot}) - 0.011 0.042 0.033 -0.011 0.034 0.025 0.070 \log(\hbox{$M_{\rm dust}$}/\hbox{$M_{\odot}$}) Mass_Dust 0.069 0.058 0.049 0.034 0.101 0.022 0.085 WAVES-like ( z\leq 0.27 ) Subsample \log(\hbox{$M_{\ast}$}/\hbox{$M_{\odot}$}) MassType_Star -0.125 0.072 0.066 -0.167 -0.078 0.017 0.155 \log(\rm{SFR}/\hbox{$M_{\odot}$}\,\rm{yr}^{-1}) StarFormationRate 0.082 0.146 0.118 -0.001 0.158 0.038 0.355 \log(\rm{sSFR}/\rm{yr}^{-1}) - 0.209 0.167 0.141 0.106 0.298 0.033 0.402 \log(\hbox{age${}_{M}$}/\rm{yr}) InitialMassWeightedStellarAge -0.219 0.091 0.085 -0.277 -0.163 0.017 0.330 \log(\hbox{$L_{\rm dust}$}/L_{\odot}) - 0.279 0.238 0.211 0.139 0.424 0.027 0.583 \log(\hbox{$M_{\rm dust}$}/\hbox{$M_{\odot}$}) Mass_Dust -0.140 0.314 0.293 -0.325 0.073 0.021 1.150 Table 4: The six parameters we investigated to determine the accuracy of the magphys model for our two additional samples discussed in Section 4 : the adjusted age prior sample at redshift z=0.10 , and the WAVES-like sample at redshift z\leq 0.27 . The columns show parameter names and units, parameter names used in the eagle database (if they are provided), median offsets, standard deviations, NMAD, Q1 offsets, Q3 offsets, outlier fractions, and median error bars of the recovered magphys values when compared to the true eagle values. Here we compare our results for the more realistic, WAVES-like sample of mock galaxies. The comparisons between the true eagle parameters and the parameters recovered by magphys are shown in Fig. 7 , with the corresponding statistical metrics listed in Table 4 . As for the idealised samples, we apply the same emissivity correction factor to the dust masses. In addition, we adopt a restricted age prior (discussed in Sect. 4 ) in order to minimise systematic stellar-mass offsets associated with mismatches between the magphys SFH priors and the eagle age distribution. From Fig. 7 and Table 4 , we find that stellar masses and star formation rates are the most robustly recovered parameters in the WAVES-like sample, with median offsets of -0.125 and 0.082 dex, respectively. This result is expected given that the nine photometric bands used in this analysis provide continuous coverage of the optical and near-infrared SED, which strongly constrains the integrated stellar emission and dust attenuation. This wavelength range captures contributions from both young and evolved stellar populations and helps to partially mitigate age-dust degeneracies, even in the absence of ultraviolet or infrared data. In contrast, dust luminosities and dust masses are substantially less well constrained. These parameters exhibit large median offsets, increased scatter, and broad posterior uncertainties, reflecting the lack of mid- and far-infrared photometry that directly traces dust emission. While energy balance allows the optical-to-near-infrared SED to place weak constraints on the total dust luminosity (e.g. 35 ; 36 ), these constraints are necessarily prior-dominated in the absence of infrared data. Dust mass estimates are even more uncertain, as they depend additionally on assumptions about dust temperature and emissivity (e.g. 33 ). The WAVES-like results therefore illustrate the limited ability of broad-band optical-near-infrared photometry alone to recover dust-related parameters with high accuracy. The increased uncertainty in the dust properties has secondary effects on other parameters. In particular, less well-constrained dust attenuation leads to enhanced age-dust degeneracies, which contribute to the modest systematic offsets observed in stellar masses and star formation rates for this sample. In this regime, dust luminosity is underestimated by \sim 0.28 dex on average, consistent with the lack of infrared constraints, and this propagates into the recovered stellar population properties. Specific star formation rates and mass-weighted stellar ages are also recovered with larger systematic offsets than in the idealised case, although their scatter and posterior uncertainties remain smaller than those of the dust parameters. The sSFR offset largely reflects the stellar-mass offset, while the challenges associated with recovering reliable stellar ages from broad-band photometry are discussed in detail in Sect. 4 . As the primary science goals of the WAVES survey focus on measuring stellar masses, star formation rates, and halo properties in the low-redshift Universe, the reduced accuracy of dust-related parameters in this WAVES-like analysis is not unexpected and does not compromise the survey’s core objectives. Instead, these results provide a clear demonstration of how limited wavelength coverage affects parameter recovery, and highlight the need for caution when interpreting dust properties, or stellar ages that are strongly coupled to dust attenuation, in the absence of infrared data (see also, e.g., discussion in 68 ). 4 Discussion 4.1 What causes the stellar mass offsets? Figure 8: The UKIDSS K -band luminosities and the mass-to-light ratios of 10,288 eagle galaxies at redshift z=0.10 , displayed in panels a) and b), respectively. The mass-to-light ratios are specifically the ratios of stellar mass in M \odot to K -band luminosity in L \odot . The true eagle values are on the y-axes and the recovered magphys values are on the x-axes. The diagonal black line represents the one-to-one relation. Printed in the bottom right corner is the median offset between the trend and the one-to-one relation, and the standard deviation of the trend. The results in Sect. 3 show that most parameters are recovered well by magphys . However, there is a significant systematic offset of about 0.140 dex in the recovered mass-weighted stellar ages of the galaxies in the redshift z=0.10 subsample, which may be related to offsets in the recovered stellar ages. Given that stellar mass is one of the fundamental parameters of galaxies that we expect to recover from SED fitting, this warrants further investigation. Both eagle + skirt and magphys use the stellar population synthesis models of 12 to model the stellar emission in galaxies, and they both assume the 22 initial mass function. Therefore, differences in modelling stellar populations, or in the assumed IMF, cannot be the cause of this offset. The stellar mass of a galaxy is dominated by low-mass stars which emit primarily in the near-infrared. We investigated whether magphys was accurately modelling the near-IR emission of the redshift z=0.10 subsample galaxies, and found a slight systematic offset ( -0.07 dex) between the K -band luminosities provided by the eagle database and the K -band luminosities of the best-fit magphys model, shown in Fig. 8 (a). magphys underestimates the K -band luminosities by 0.07 dex, and we note that this offset is very consistent across the sample. This translates into a similar offset in K -band mass-to-light ratio. Given that our observed stellar mass offset is even larger (Fig. 5 ), this offset is not the sole contributor to the difference in stellar mass. The mass-to-light ratios of a galaxy in the near-IR are affected by its star formation history (SFH) and secondarily affected by its dust content (e.g., 24 ). The star formation history of a galaxy describes the star formation rate of the galaxy over the length of its existence, including any bursts of star formation due to interactions with other galaxies or similar processes. It is notoriously difficult to recover the SFH of a galaxy from its broad-band SED alone, as multiple different SFH can result in similar looking SEDs ( 76 ; 77 ; 32 ) , and this is further compounded by age-dust degeneracies that affect the UV/optical colours. Spectral energy distribution models use various templates or parametric functions that attempt to replicate the complex history of star formation in a galaxy over its lifetime. For example, magphys uses a linearly rising and then exponentially declining (‘delayed-tau model’) with superimposed random bursts to generate its library of possible SFHs for input SEDs (see section 3.1.1 of 34 ). A difference in the simulated star formation histories of the eagle galaxies and the possible templates generated by magphys could be indicated by the significant discrepancies in the stellar ages, and is a plausible cause of the difference in the current stellar masses ( 21 ; 66 ) . Figure 9: The distributions of the original magphys age prior (blue), the adjusted magphys age prior (light blue), and the age values of all 10,288 eagle galaxies (yellow) from the redshift z=0.10 subsample of galaxies. The vertical black line at 10 10.1 years represents the oldest age a stellar population can be at this redshift (set by the age of the Universe). In Fig. 9 , we compare the distribution of the magphys mass-weighted age priors with the distribution of all the age values from the eagle models for the galaxies in our redshift z=0.10 subsample. The eagle mass-weighted stellar ages range almost exclusively from 10^{9.5} to 10^{10} years. This is a very small and old section of the range covered by the magphys age priors, which extend to mass-weighted ages as low as \sim 10^{7.5} years. To investigate the effect of this discrepancy on the recovered parameters, we generated a new magphys age prior adjusted to match the eagle age distribution, which is shown in light blue in Fig. 9 . Essentially, we exclude magphys models with ages younger than 10^{9.5} years from the prior distribution. We then re-fit the redshift z=0.10 subsample with this new age prior and compare the recovered stellar masses and mass-weighted ages with their true values in Fig. 10 . We note that the other galaxy parameters (SFR, sSFR, M_{\rm dust} , L_{\rm dust} ) remain well-constrained with low offsets, and for completeness we show the new comparisons for these parameters in Appendix A . The statistics of these comparisons are also displayed in Table 4 . Fig. 10 shows that we recover the stellar masses of the eagle mock galaxies with much smaller offsets when using the adjusted magphys age prior. The median offset value of the stellar mass offset has been reduced from 0.140 dex in Fig. 5 (a) to \mathbf{0.029} dex in Fig. 10 (a). A similar reduction is found in the mass-weighted stellar ages, with the median offset value changing from 0.261 dex in Fig. 6 (a) to \mathbf{0.081} dex in Fig. 10 (b). Figure 10: Comparisons of the stellar masses and mass-weighted stellar ages of the 10,288 eagle galaxies from the redshift z=0.10 subsample, after removing younger ages from the magphys age distribution, in panels a) and b), respectively. The diagonal black line represents the one-to-one relation. Printed in the bottom right corners of the panels are the median offsets between the trends and the one-to-one relation, and the standard deviations of the trends. Printed in the top left corners of the panels are the median error bars of the recovered values, defined as the median difference between the 16 th and 84 th percentiles of each recovered value. These results support our hypothesis that the offsets of the stellar ages and the stellar masses found when recovering the physical parameters of eagle galaxies with magphys are caused by the differences between the histories of the eagle galaxies and the SFH prior of the magphys model. It is interesting that even though magphys is able to fit the SEDs of the eagle galaxies well, the choice of SFH prior has a significant effect on the recovery of the stellar masses of those galaxies. We note that similar results were found by 82 in a test fitting the UV-to-near-IR SEDs of z\sim 2 mock galaxies from the semi-analytic model. They concluded that the stellar masses of galaxies underestimated because the mock galaxy SFHs were rising, whereas their SED model assumed an exponentially-declining form. When using rising SFHs, the offsets in stellar masses decreased. Our results beg the question: are the magphys SFH priors inappropriate to model low-redshift galaxies? While they are certainly ‘too young’ for the subset of eagle redshift z=0.10 galaxies presented here, we must be cautious about taking these results as a sign that the magphys priors need fine-tuning. This is because there are indications that the eagle stellar ages may not be representative of the stellar ages of galaxies in the local Universe. For example, 101 compared galaxies at z\simeq 0 from multiple simulations, including eagle , to galaxies from integral field unit (IFU) surveys such as the SAMI (Sydney-AAO Multi-object Integrated field spectrography) Galaxy Survey ( 31 ; 13 ) , the ATLAS 3D Survey ( 19 ) , the SAURON Survey ( 41 ) , the CALIFA (Calar Alto Legacy Integral Field Area) Survey ( 91 ) , and the MASSIVE Survey ( 69 ) . They find that within the same stellar mass range ( \log(\hbox{$M_{\ast}$}/\hbox{$M_{\odot}$})<11.5 ) the simulated eagle galaxies tend to have much older luminosity-weighted ages than the observed galaxies of their sample. They find no eagle galaxies with ages <10^{9.4} yr, while the sample of observed galaxies extends to ages below 10^{8.6} yr. Figure 11: Comparison of the stellar age distribution of eagle and SDSS DR4 galaxies. The eagle galaxies are shown in yellow, and the SDSS galaxies are coloured by the number of galaxies per cell as seen in the colour bar. The stellar mass lower limit of the eagle galaxies at 10 9.5 M \odot is clear when compared to the lower limit of the SDSS galaxies at 10 7.5 M \odot . We further investigate how the average stellar ages of our eagle galaxies compared to those of observed galaxies at redshift z=0.10 by analysing data from the Sloan Digital Sky Survey (SDSS). We compared the distribution of average stellar ages of 567,486 SDSS galaxies that lie between 0.005