EDP Sciences logo
Subscriber Authentication Point
Search
Menu
Home All issues Volume 699 (July 2025) A&A, 699 (2025) A336 Full HTML
Open Access
Issue
A&A
Volume 699, July 2025
Article Number A336
Number of page(s) 17
Section Extragalactic astronomy
DOI https://doi.org/10.1051/0004-6361/202554231
Published online 23 July 2025

© The Authors 2025

Licence Creative CommonsOpen Access article, published by EDP Sciences, under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

This article is published in open access under the Subscribe to Open model. Subscribe to A&A to support open access publication.

1. Introduction

After the big bang, nucleosynthesis started within the first stellar population formed in the Universe: that is, population III stars (pop. III) and, shortly after, pop. II stars. Type II supernovae (SNae II) expelled metals much earlier than asymptotic giant branch (AGB) stars in the early Universe (Valiante et al. 2009; Hirashita et al. 2014; Dell’Agli et al. 2019, and Walter et al. 2020), which then led to the formation of the first dust grains via the coalescence of these metals (e.g. Schneider & Maiolino 2024). Dust affects ultraviolet (UV) and optical emissions through dust attenuation and reddening. Dust infrared (IR) emission is a major cooling agent. This energy warms dust and is re-emitted at IR and sub-millimeter (sub-mm) wavelengths (Burgarella et al. 2005; Małek et al. 2018, and Pozzi et al. 2021). Dust plays a critical role in the formation of low-mass stars by facilitating several key processes in the interstellar medium (ISM). Molecular hydrogen (H2), which is essential for cloud collapse and subsequent star formation, does not form efficiently in the gas phase under typical ISM conditions. Instead, H2 formation is catalysed on the surfaces of dust grains, making dust indispensable to initiating star formation even at warm temperatures (Grieco et al. 2023). Once formed, these grains can act as seeds for further grain growth in the ISM, enhancing the dust mass available (Zhukovska et al. 2018 and Asano et al. 2013). Furthermore, collisions between gas and dust grains enable efficient gas cooling, particularly at high densities (nH∼1012 cm−2), which promotes fragmentation of the cloud and the formation of low-mass stars. These low-mass stars may represent a transition population between pop. III and pop. II stars.

One of the main results from the JWST's first years of observation is an unpredicted excess of UV-luminous galaxies at z > 10 compared to HST-calibrated models (Finkelstein et al. 2023, 2024; Naidu et al. 2022, and Casey et al. 2024). The galaxies present far-UV (FUV) absolute magnitudes, –21 ≲ MUV ≲ –19; very blue FUV spectral slopes, βFUV ≲ –2.2; small effective radii, re ∼ 200–400 pc, and stellar masses, Mstar ∼ 109 M⊙ (Atek et al. 2023 and Ferrara et al. 2025) that are similar to ours, especially for the highest redshift galaxies.

Several potential causes have been proposed. We cannot present an exhaustive inspection in this paper, and we suggest following some of the tracks opened in the next paragraphs for a full and detailed review.

A first hypothesis (Feldmann et al. 2025), is proposed whereby relatively high star formation efficiency (SFE) in the early Universe is a natural outcome of the baryonic processes encoded in the FIRE-2 model (Hopkins et al. 2018). This is because the shallower slope of their SFE-Mhalo relation at 9<log10(Mhalo/M⊙)<11 leads to an increase in the contribution from the more numerous lower mass haloes, and thus an increase in the observed abundances of bright galaxies at z > 10.

Another proposed hypothesis is that this higher SFE could be due to feedback-free starbursts (FFBs), where the SFE could reach a maximum SFE = 0.2–1.0 in the FFB regime (Li et al. 2024), to the formation of pop. III stars (e.g. Yung et al. 2024 or dark stars with masses ≳ 103 M⊙), which could be fuelled by heating from dark matter in the first dark-matter halos or minihalos (Ilie et al. 2023 and Lei et al. 2025). The effect could be due to a top-heavy initial mass function (IMF): an increase in the characteristic stellar mass of a top-heavy IMF would add up massive stars, which in turn would produce more UV light (e.g. Zackrisson et al. 2011; Harikane et al. 2023; Hutter et al. 2025, and Jeong et al. 2025).

It could also be due to an increased stochasticity exposed through dispersion in the relation between galaxy UV magnitude (MUV) and halo mass: bursty star formation histories (SFHs) have been measured (e.g. Cole et al. 2025) via the scatter in the star-forming main sequence (MS). However, while such stochasticity in star formation can certainly help, stochasticity alone might not be enough (Yung et al. 2024 and Finkelstein et al. 2024).

Finally, an origin related to dust, either as the only process or combined with another is certainly quite compelling. This could be a low dust attenuation and/or a specific dust-star geometries directly leading to unobscured young stars.

In the last decade, we have detected dusty galaxies at z > 4 (Vieira et al. 2013; Hezaveh et al. 2015; Watson et al. 2015; Laporte et al. 2017; Fudamoto et al. 2017; Strandet et al. 2017; Tamura et al. 2019; Sommovigo et al. 2022a; Algera et al. 2024a, b; Witstok et al. 2023; Zavala et al. 2023, and Valentino et al. 2024) with large dust masses (Pozzi et al. 2021 and Akins et al. 2023) that cannot be explained by models. SNae and AGB stars could scarcely be at the origin of such a large dust mass, especially if we account for the reverse shock resulting from the expanding SN blast wave in the ISM (Leśniewska & Michałowski 2019). To explain such a large dust mass, we could assume an important role of dust mass growth in the ISM.

Models of dust formation, both semi-analytical (e.g. Popping et al. 2017; Vijayan et al. 2019; Triani et al. 2020; Dayal et al. 2022, and Mauerhofer & Dayal 2023) and cosmological (e.g. Graziani et al. 2020; Lewis et al. 2023; Di Cesare et al. 2023, and Choban et al. 2024) predict a critical metallicity in the ISM above which grain growth via gas–dust accretion becomes efficient, marking the transition from dust production dominated by stellar sources (supernovae and AGB stars) to accretion-driven dust growth in the ISM. While this transition is supported by local-Universe observations (Rémy-Ruyer et al. 2014 and De Vis et al. 2019, it remains undetected at high redshift.

Ferrara et al. (2022) suggested that dust could have been efficiently ejected during the very first phases of galaxy build-up. This hypothesis is supported by theoretical predictions that high-redshift galaxies are extremely bursty and very small (e.g. Sun et al. 2023 and Choban et al. 2024). In this case, we would only detect the almost dust-free objects that remain. However, both star or ISM dust geometry and galactic dust ejections would produce the same low dust attenuations and thus bright UV luminosities. Other discriminant observables such as the gas-mass fraction and the metallicity should be explored to study this possible degeneracy.

This work derives new constraints on the ISM at high redshift from the JWST-CEERS project1, which features NIRCAM and NIRSpec, as well as ancillary data from SCUBA-2 (Zavala et al. 2018) and NOEMA (Fudamoto et al. 2017), with the help of a new version of the CIGALE code that accepts spectrophotometric data. More technical details on the new version of CIGALE are provided in the Appendix A.

2. Observations

2.1. NIRSpec prism spectroscopic sample

2.1.1. Origin of the sample

We have 1337 spectroscopic observations with NIRSpec (Arrabal Haro et al. 2023) from CEERS. We used 634 of these NIRSpec observations carried out with the prism configuration. The spectroscopic targets are selected on the basis of CANDELS HST imaging (Grogin et al. 2011 and Koekemoer et al. 2011) on various non-homogeneous criteria, which might introduce a bias in the selection. CEERS's NIRCam observations detect 101 808 objects photometrically (CEERS_v0.51.4, Bagley et al. 2023). Whenever possible, we combined spectroscopic data (Birkmann et al. 2022) with photometric data by cross-matching the coordinates within 0.2 arcsec. However, because there is no complete overlap between CEERS NIRSpec and NIRCam observations, some NIRSpec fields are in areas where we have no NIRCam imaging. For these, we only used the spectroscopic data. After fitting the prism spectra (with and without NIRCam data), we checked the quality of the spectroscopic redshifts for objects with z spec > 4.0. We sorted the redshifts into four quality classes from qz = 0 (no doubts on redshift) to qz = 4 (wrong or unconfirmed redshift). We kept 173 objects with NIRSpec observations, for which qz = 0 (all modelled lines match the observed spectrum) or qz = 1 (some fainter lines are not in excellent agreement with the models). Objects with qz = 2 present a continuum that could be in agreement with the derived redshift, but without any positively identified lines, and objects with qz = 3 only hint at a low signal-to-noise ratio (S/N) for the estimated redshift. We flagged a sample of six possible AGNs based on the literature (Kocevski et al. 2023; Larson et al. 2023, and Harikane et al. 2023). These AGNs are listed in Table 1 and shown with crosses in the plots. From this analysis, the distribution of redshifts is shown in Fig. 1. Most galaxies are clustered in the z = 4.0–8.0 range, with a small but important tail extending to z ≲ 12.

thumbnail Fig. 1.

Distribution of spectroscopic redshifts derived by fitting spectrophotometric data with CIGALE. We kept 173 objects with the most robust redshifts. Galaxies are mainly found in two sub-populations at 4 ≲ z ≲ 6 and 5 ≲ z ≲ 6, with a tail extending to z ≲ 12.

Table 1.

Extract from Harikane et al. (2023): possible AGNs in the analysed sample.

2.1.2. Analysis of spectrophotometric data

We used all data with a S/N of S/N > 1.0 per pixel in the spectrum and for each NIRCam photometric band. All other measures were set as upper limits in the fit. We utilised a new version of CIGALE that accepts both photometric and spectroscopic data (see Appendix A), and we define this type of mixed data as spectrophotometric energy distributions (SPEDs) to clarify the difference between these and traditional spectral energy distributions (SEDs). The priors used in the fits are listed in Table B.1, and a sample of the fits is shown in Figs. B.1–B.6.

Two SFHs were used to test the stability of the results: a delayed-plus-burst one and a periodic one. The delayed SFH assumes that star formation is active over a few tens to hundreds of mega-years with SFR(t) ∝  t τ 2 × exp − t / τ $ \frac {t}{\tau ^2} \times \exp ^{-t/\tau } $ followed by a final burst with various possible ages. Various other types of SFHs could be used, but determining the SFH in the early Universe is difficult for any SED modelling method (Lower et al. 2020). Moreover, Leja et al. (2019), Iyer et al. (2019) and Tacchella et al. (2023) insisted on the fact that the priors chosen for the fit are the primary drivers, ahead of the type of SFH, used to recover the physical parameters. The number of priors thus sets strong constraints on the ability to successfully run fitting codes, especially for large samples. The speed of CIGALE (Burgarella et al. 2024) allows it to explore various sets of priors with several 108 models in a reasonable time for thousands of spectrophotometric objects. For the alternate SFH, a periodic SFH is chosen because it is conceptually different from the delayed-plus-burst SFH: it does not assume any kind of continuous SFH. Instead, a series of bursts, separated by regular quiescent periods, are used (see Appendix C, which shows that this periodic SFH has a small impact on the results presented in this paper).

CIGALE estimates line fluxes through a comparison with its Cloudy-derived nebular models. To be sure we are on safe ground, we validated the line fluxes measured by CIGALE with those estimated via other methods: in Figs. 2 and 3, we checked that the flux of the emission lines measured by CIGALE are consistent with first fluxes estimated with the LiMe software (Fernández et al. 2024); and, second, we performed a fit on the sub-sample where both prism and grating NIRSpec spectra were available. For this second measurement, we fitted the lines with lmfit (based on the Levenberg-Marquardt method). lmfit provided us with tools to perform non-linear optimisation and curve fitting in Python. Practically speaking, we fitted the lines per group of three (e.g. Hβ + [OIII] or [NII] + Hα) assuming a model that is the sum of three Gaussian distributions for the emission lines, plus a line to fit the continuum. The comparison is good up to line fluxes of about 3σ of the local background.

thumbnail Fig. 2.

Comparison of CIGALE fluxes to those measured with LiMe (Fernández et al. 2024). The full results on the CEERS catalogue redshifts and line measurements will be discussed in Arrabal Haro et al. (in preparation). The points corresponding to each emission line are colour-coded to better identify each species (see inside caption). AGN lines are identified by crosses. The most ultraviolet lines (CIV λ1549 Å and HeII λ1641 Å in the lower part of the plot) are not in agreement with those estimated by LiMe as shown by the offset from the solid 1-to-1 line. Instead, we find that other lines are systematically different by about about 20% (dashed line). The observed difference might find an origin in the subtraction of the continuum.

thumbnail Fig. 3.

Comparison of fluxes computed by CIGALE (Fluxprism) with Gaussian fitting of the lines observed in the grating configuration (Fluxgrating) for objects observed in both configurations. The points corresponding to each emission line are colour-coded to better identify each species (see inside caption). The diagonal line shows the 1-to-1 relation (Fluxprism = Fluxgrating), while the dashed lines present offsets (Fluxprism = 0.5 Fluxgrating) and (fluxprism = 2.0 Fluxgrating). The horizontal and vertical dashed lines show the level of the 3σ background around the [NII] + Hα lines.

3. The properties of the galaxy sample

In this section, we define how we selected a specific population of 49 galaxies that have extremely low dust attenuation (GELDAs). We observationally defined GELDAs using the following criteria:

  • The FUV dust attenuation AFUV = 0.0 within 2σA_FUV.

  • There is a stellar mass of Mstar < 109 M⊙.

3.1. Star formation rate and stellar mass

Figure 4 presents the histogram of the stellar masses derived by CIGALE for the present sample of galaxies. The structure of the histogram as a function of the UV slope βFUV confirms the relation between stellar mass and dust attenuation at high and ultra-high redshifts, as shown by several papers (Bogdanoska & Burgarella 2020; Weibel et al. 2024, and Bouwens et al. 2016).

thumbnail Fig. 4.

Distribution of stellar masses derived by fitting spectrophotometric energy distributions of our galaxy sample. In this plot, we only show galaxies with βFUV derived from CIGALE's fits. The range of stellar masses reaches Mstar values as low as a few 107 M⊙. The colour-coding shows that the least massive galaxies clearly have bluer UV slopes (βFUV≤−2.2), while the most massive galaxies have redder slopes (−1.5<βFUV), as expected from the dependence of UV dust attenuation on stellar mass (see e.g. Bogdanoska & Burgarella 2020; Weibel et al. 2024; Bouwens et al. 2016). The vertical numbers show the mean stellar mass for each sample.

Figure 5 shows the location of the MS in the SFR versus Mstar diagram. Because the redshift range is quite large, we might not expect all galaxies to follow a tight sequence. However, since most of the sample is in the 4 < z < 6 redshift range (Fig. 1), we do observe such a sequence, which is structured by the UV slope βFUV. We do not see any noticeable differences between UV slopes derived by directly fitting the spectra and those derived by fitting the SPEDs with CIGALE. In Fig. 5, we present a comparison of the location of our sample with the best-fit MS from Speagle et al. (2014). The fits from Speagle et al. (2014) show the well-known strong increase in SFR at low redshifts (z ≲ 8), followed by a gradual lower increase at higher redshifts up to z = 12. Our sample of objects are mainly in the first redshift bin (4 ≲ z ≲ 8; see Fig. 1). They are MS galaxies. However, part of the sample lies above the highest MS from Speagle et al. (2014), especially the reddest. They likely belong to a star-busting class of galaxies. AGNs are preferentially found at high masses and high SFRs.

thumbnail Fig. 5.

Star formation rate as function of stellar mass for our galaxy sample. For some galaxies, the observed spectra are not good enough to provide a reliable fit of the spectrum (f λ ∝ λ FUV β $ {}_{\lambda } \propto \lambda ^\beta _{{\mathrm {FUV}}} $) for the UV slope. For those galaxies, we rely on CIGALE's fits. Here, the larger symbols correspond to UV slopes spectroscopically estimated from the observed data, while the smaller symbols are from the CIGALE fits. Blue, green, and red symbols, respectively, show galaxies with blue, intermediate, and red slopes (see inside caption). Whatever the origin of the UV slopes, the lower mass galaxies present a bluer UV slope, while the more massive ones have are redder. In this plot, we superimpose the best-fit MS from Speagle et al. (2014) in the range z = 4–12, colour-coded from magenta to dark red (bottom to top).

In Figure 6, we show the far-UV dust attenuations, AFUV, versus log10(Mstar), where we again note a large dispersion at low Mstar spanning more than two decades. The stellar mass is not the only parameter acting, as both low and relatively high AFUV lie in the same stellar-mass range. We do not observe any clustering of low-redshift galaxies (z ≲ 8.8) in the lower part of the plot. This shows that these low-redshift galaxies could have a wide range of dust attenuation. On the other hand, five out of six of the highest redshift galaxies (z ≳ 8.8) are found in this part of Fig. 6, and they are therefore GELDAs. This would suggest that GELDAs become dominant in the early Universe.

thumbnail Fig. 6.

Far-UV dust attenuation AFUV as a function of stellar mass, log10(Mstar). The symbols are colour-coded and sized with the redshift. The red objects correspond to the highest redshift bin defined in this paper (z > 8.8). The majority of highest-redshift GELDAs (5 out of 6) are found in the low-AFUV part of the plot. This sample of galaxies at z > 8.8 is small, but it might confirm an expected decreasing evolution of the dust attenuation with the redshift. The dashed line is the ‘consensus’ law from Bouwens et al. (2016), while the continuous line is the best linear fit (AFUV vs log10(Mstar)) to the present data. GELDAs are shown with an upside-down Y, and AGNs are shown with crosses.

The ‘consensus’ law between AFUV and Mstar estimated at the lower redshift (z ∼ 2–3) by Bouwens et al. (2016) does not pass through the present data. Bogdanoska & Burgarella (2020) found that this consensus law does not appear to be valid at high redshifts. They also suggested that the low-stellar-mass galaxies exhibit a large scatter in AFUV and proposed an evolution of this AFUV and Mstar relation with the redshift. However, the sample available in Bogdanoska & Burgarella (2020) could hardly reach log10(Mstar) ≲ 9.0. The JWST's sensitivity to much fainter flux allows one to reach galaxies at much lower stellar masses, and potentially also permits one to detect this new population of GELDAs.

3.2. Metal and gas masses of the galaxy sample

In CIGALE, the stellar Zstar and nebular Zgas metallicities have different priors. This is most useful when spectroscopic data are available because they allow us to separately constrain both metallicities by fitting the spectra, including lines. For this specific estimation process, CIGALE makes use of the new nebular models described in Theulé et al. (2024) that can have excitation parameters up to logU = –1.0 and with a wide range of nebular metallicities and electronic densities (nH). The metallicity Zgas derived by CIGALE can be converted into 12 + log10(O/H) where O/H is the oxygen abundance of the gas by using their Table 1 (correspondence among ξ0, the interstellar gas metallicities, and the stellar metallicities). ξ0 is defined on oxygen abundance: ξ0 = (O/H)/(O/H)GC, where (O/H)GC = 5.76 × 10−4, and GC is the so-called local Galactic concordance (Theulé et al. 2024 follows Nicholls et al. 2017). From this, Eq. (1) links total metallicity to oxygen abundance:

12 + log 10 ( O / H ) ) = log 10 ( Z gas ) + 10.410 . $$ 12\,+\,\log _{10}({\mathrm {O/H}})) = \log _{10}({\mathrm {Z}}_{\mathrm {gas}}) + 10.410. $$(1)

In Fig. 7, we compare the CIGALE metallicities estimated for the same CEERS galaxies by Nakajima et al. (2022, 2023), and Sanders et al. (2024). Nakajima et al. (2023) measured emission-line fluxes for 135 galaxies (115 in CEERS). For ten of these galaxies, they determined their electron temperatures with [O III] λ4363 Å lines, in a way similar to lower redshift star-forming galaxies, and they derived the metallicities using a direct method. They finally estimated metallicities for their entire sample of JWST-observed galaxies with strong lines using their previous metallicity calibration (Nakajima et al. 2022), based on the direct-method measurements. Our sample in common with Nakajima et al. (2023) amounts to 90 galaxies. Sanders et al. (2024) also combined JWST measurements with [O III] λ4363 Å auroral line detections from JWST-NIRSpec and from ground-based spectroscopy to derive electron temperature (Te) and direct-method oxygen abundances on a combined sample of 12 star-forming galaxies at z = 1.4–8.7. Our sample shares three of them, for which we derive metallicities with CIGALE. Finally, Nakajima et al. (2023) and Sanders et al. (2024) have three common objects. Our fitting method is different, as the total spectrophotometric fits allow one to consistently constrain the metallicities (Zstar and Zgas) by only selecting models that agree with all information brought by observations; that is, the continuum and lines together.

thumbnail Fig. 7.

Difference between metallicity estimates (12 + log10(O/H) defined in Eq. (1)): a comparison of the metallicities measured for some of our galaxies in the literature (Nakajima et al. 2022, 2023; Sanders et al. 2024) via auroral lines, with our own estimates. At relatively high metallicities, above 12 + log10(O/H) ∼ 7.4–7.6 (that is Z/Zgas 0.1–11% Z⊙), the differences remain within σ(12 + log10(OH))<0.05. This is about the same dispersion between other metallicity estimates, although on a smaller sample. However, there is a disagreement at lower metallicities, and CIGALE's 12 + log10(O/H) present an offset that might be systematic or increasing with lower metallicities by about σ (log10(O/H)) < 0.10. But the direct method is not well calibrated at log10(O/H) < 7.4–7.6 because there are less than 5 objects with auroral lines at such low metallicities estimated via the direct-method oxygen abundances.

The metallicities estimated by CIGALE show a systematic or gradually increasing discrepancy at log10(O/H)<7.4−7.6, with Nakajima et al. (2023) metallicities lower by about σ (log10(O/H))<0.10. However, we note that the direct method is not well calibrated below log10(O/H)<7.4−7.6, because there are less than 5 objects with auroral lines at such low metallicities estimated via the direct-method oxygen abundances. The true value becomes quite uncertain until we obtain a better calibration. Fig. 8 shows that, individually and globally, the metallicity estimates can vary depending on the method used to derive metallicities from the lines. The CIGALE metallicities estimated by fitting the entire spectrum are found close to most values from Nakajima et al. (2023) and Sanders et al. (2024), and especially very close to the R23 index, which is found to be the most reliable among various metallicity indicators over the wider range (Nakajima et al. 2022). However, the line ratios involving nitrogen lines (N2 and O3N2 from Nakajima et al. 2023) lead to much lower values by about 1 dex. Nakajima et al. (2022) found a scatter as large as Δlog10(O/H)∼0.4 dex in the relation for metallicities derived using the N2 index. This is especially true at low metallicities, as for our galaxies. They suggested that this might be associated with line ratios that use single-ionised, low-ionisation lines such as [NII]. We conclude that care should be taken when using only one of these indices. However, CIGALE provides us with a safe and reliable method for estimating metallicities, at least in the present range.

thumbnail Fig. 8.

Direct comparison of metallicities (12 + log10(O/H) defined in Eq. (1). Left: this panel shows the metallicities derived using various published calibrations as a function of the mean metallicities computed from each measured galaxy. The dashed black line is the 1-to-1 relation. For a given metallicity on an x-axis, we could get a very wide range on the y-axis depending on which line ratio is used. The black dots are the metallicities from CIGALE. The red-orange ones from Nakajima et al. (2023) and the blue-green one from Sanders et al. (2024). The color coding is given in the legend inside the plot. Right: this panel presents the same information with a rolling average. Most of the metallicity values agree within about ±0.5 dex and the metallicities from CIGALE are approximately in the middle range. However, we notice that the ones involving the nitrogen lines lead to much lower metallicities (see text for more details.

Figure 9 presents the specific metal masses (MZ/Mstar) for the present sample as a function of the specific SFR (SFR/Mstar). The metal mass, MZ, is computed from Eq. (2) (Heintz et al. 2023a):

thumbnail Fig. 9.

Evolution of metal-to-stellar mass ratio as function of specific star formation rate in this metal formation rate diagram (MFRD) presented in the same way we defined the dust formation rate diagram (DFRD) showing how the specific dust mass (Mdust/Mstar) changes with the specific star formation rate (SFR/Mstar) in Burgarella et al. (2022). This type of diagram is normalised to Mstar and allows a better comparison between galaxies. The symbols are colour-coded with the redshift. The pink shaded area presents the 2-σ confidence interval for the distribution in MZ/Mstar, and the solid black line shows the rolling average. We observe a decreasing trend with decreasing sSFR. The highest redshift galaxies are located in the lowest part of the plot.

M Z = M gas × 10 12 + log 10 ( O / H ) − 8.69 × Z ⊙ . $$ M_{\mathrm {Z}} = M_{\mathrm {gas}} \times 10^{12\,+\,\log _{10} ({\mathrm {O/H}})-8.69} \times Z_\odot . $$(2)

In Eq. (2), the gas mass, Mgas, and the oxygen abundance, 12 + log10(O/H), are estimated via the spectrophotometric fitting and from Eq. (1). To estimate Mgas, we need to estimate the molecular mass Mmolgas from Eq. (3) (Tacconi et al. 2020):

log 10 ( M molgas ) = 0.06 − 3.33 × [ log 10 ( 1 + z ) − 0.65 ] 2 + 0.51 × log 10 ( sSFR / sSFR ( MS , z , M star ) − 0.41 × [ log 10 ( M star ) − 10.7 ] × M star , $$ \begin{aligned}\log _{10} ({\mathrm {M}}_{\mathrm {molgas}}) &= 0.06-3.33 \times [\log _{10}(1 + z) - 0.65]^2\\ & \quad + 0.51 \times \log _{10}({\mathrm {sSFR/sSFR}}({\mathrm {MS}}, {z}, {\mathrm {M}}_{\mathrm {star}})\\ & \quad - 0.41 \times [\log _{10}({\mathrm {M}}_{\mathrm {star}}) - 10.7] \times {\mathrm {M}}_{\mathrm {star}}, \end{aligned} $$(3)

where the specific instantaneous SFR, sSFR = SFRinst/Mstar, is derived from the spectrophotometric fitting, while the reference sSFR for the MS as a function of the stellar-mass Mstar and the redshift, sSFR(MS, z, Mstar), is from Speagle et al. (2014), assuming the so-called Bluer w/ high-z obs MS in their Table 9 (Eq. (4)). The MS is modelled up to z ∼ 6, while our sample reaches z = 11.4. However, even if the scatter in the MS is quite large, studies suggest that it should not show any strong evolution to z ∼ 12 (Cole et al. 2025 and Chakraborty et al. 2024):

log 10 ( SFR ( M star , Age Universe ) ) = [ ( 0.73 − 0.027 × Age Universe ) × log 10 ( M star ) − ( 5.42 + 0.42 × Age Universe ) ] − log 10 ( M star ) . $$ \begin{aligned}\log _{10} ({\mathrm {SFR}} ({\mathrm {M}}_{\mathrm {star}}, {\mathrm {Age}}_{\mathrm {Universe}})) & = [(0.73 - 0.027 \times {\mathrm {Age}}_{\mathrm {Universe}})\\ & \quad \times \log _{10} ({\mathrm {M}}_{\mathrm {star}}) - (5.42 + 0.42 \\ & \quad \times {\mathrm {Age}}_{\mathrm {Universe}})] - \log _{10} ({\mathrm {M}}_{\mathrm {star}}). \end{aligned} $$(4)

We also need to add the contribution from the atomic gas Matomgas to Mmolgas The atomic-to-molecular mass ratio Matomgas/Mmolgas was estimated by Chowdhury et al. (2022) for star-forming galaxies at z ∼ 0, z ∼ 1.0, and z ∼ 1.3, with values in the range of 4 ± 2 for galaxies with Mstar>1010 M⊙. To account for galaxies with <1010 M⊙ in their statistics, they assumed that the ratios Matomgas to Mmolgas are systematically higher by a factor of about five for all galaxies with Mstar < 1010 M⊙. In this case, the values obtained would increase Mmolgas at z ∼ 1.3 by a factor of approximately two, giving a ratio of Matomgas/Mmolgas = 2.5 for the highest redshifts. At higher redshifts (0.01 < z < 6.4), there is no significant redshift evolution of the Matomgas/Mmolgas ratio (Messias et al. 2024), which is about 1–3. At z = 8.496, the gas and stellar contents of a metal-poor galaxy were studied with the JWST and ALMA (Heintz et al. 2023b). From this analysis, they infer Mmolgas = (3.0–5.0) × 108 M⊙, corresponding to 40% ± 10% of Mgas for their object, which leads to Matomgas/Mmolgas = 1.5 1.3 2.0 $ {}^{2.0}_{1.3} $. Given the redshift of our objects, in this work we assumed Matomgas/Mmolgas = 2.0 1.3 2.5 $ {}^{2.5}_{1.3} $.

3.3. Dust masses

Dust masses (Mdust) are very important for collecting diagnostics on the origin of dust in galaxies because they are directly related to the dust-building rate, which in turn can provide us with information on the origin of dust grains (e.g. Leśniewska & Michałowski 2019; Nanni et al. 2020; Burgarella et al. 2022). To help estimate the dust masses for our objects, we used deep 450 and 850 μm SCUBA-2 and NOEMA-1.1 mm observations. Both are cross-correlated with JWST's coordinates. The former have a mean depth of σ450 = 1.9 and σ480 = 0.46 mJy beam−1, and the angular resolution is θFWHM≈8 arcsec at 450 μm, and θFWHM≈14.5 arcsec at 850 μm. For NOEMA, the rms is σ1.1 mm = 0.10 mJy beam−1, and the beam size is 1.35 × 0.85 arcsec2. Although the sensitivity of these observations is typically lower than that required to detect most of our galaxies, the inferred upper limits are useful for putting constraints on the total IR luminosities and dust masses. Similarly, the angular resolutions are much larger than those of the JWST (Zavala et al. 2023). Some of the associations might therefore be wrong. However, these sub-millimetre data rule out any strong, lower-redshift, far-IR emitters associated with the objects in our sample.

The far-IR information for this sample is limited, and we cannot directly derive any information on the dust emission SED shape. However, the ALMA-ALPINE sample, for example, Béthermin et al. 2020; Pozzi et al. 2021, and Sommovigo et al. 2022a, and the ALMA-REBELS samples (e.g. Inami et al. 2022; Sommovigo et al. 2022b; Algera et al. 2024a, b) present physical properties including stellar masses of log10(Mstar) ∼ 10 and redshifts of 4.5 < z < 7.7, which are similar to ours (Burgarella et al. 2022 and Nanni et al. 2020). Due to the similarities of the samples, we assume that we can make use of the same Draine et al. (2014) best-fit model (see Table B.1) identified in Burgarella et al. (2022) and Nanni et al. (2020). We note that this ALMA-ALPINE model corresponds to a dust temperature of Tdust = 54.1 ± 6.7 K, assuming an optically thin modified black body, which is in good agreement with Sommovigo et al. (2022a), which found an average value < Tdust> = 48 ± 8 K and Mdust in the (0.5–25.1) × 107 M⊙ range for ALPINE. For ALMA-REBELS, the median Tdust is in the 39–58 K range, and the median dust masses are estimated in the (0.9–3.6) × 107 M⊙ (Sommovigo et al. 2022b) range. Sommovigo et al. (2022b) also predicted that dust masses can be produced by SNae alone for 85% of the REBELS sample. We note that Algera et al. (2024a, b) found lower dust temperatures (Tdust = 30–35 K) for two objects in the REBELS sample. Furthermore, Sommovigo et al. (2022a) predicted that more metal-poor, high-z galaxies could have higher temperatures because of their lower dust content, while the objectives studied in Algera et al. (2024a, b) are metal-rich. However, the dust model is assumed to be the same in our work for all of our objects. Globally modifying our model would cause an offset of all the dust masses, but not the observed relative difference between GELDAs and non-GELDAs.

Because we used the above single-dust emission model (Draine et al. 2014), we did not derive any shape for the IR emission in this work. The shape of the IR spectrum is fixed by the ALMA-ALPINE sample at 4.5 < z < 6.2, and the IR luminosity was estimated assuming the energy balance concept; Mdust is constrained by the amount of dust attenuation and by the main observables that define this dust attenuation. Information on the amount of dust attenuation comes from the line ratios, especially Hα/Hβ when available; the UV slope βFUV; and from any available IR or sub-millimetre data (Figs. 10 and 11).

thumbnail Fig. 10.

Correlation of estimated dust mass, Mdust, with dust attenuation, A(Hα). The most relevant parameter to predict the dust mass is the dust attenuation A(Hα) derived from the Hα/Hβ Balmer decrement. The correlation with A(Hα) alone accounts for 76% of the variation in Mdust. The spectral information brought by NIRSpec is thus fundamental to estimating the dust masses. Blue, green, and red symbols, respectively, mean Mdust≤105 M⊙, 105 < Mdust≤106 M⊙, and Mdust > 106, while magenta circles show GELDAs.

thumbnail Fig. 11.

Same colours as in Fig. 10. Correlation of the estimated dust mass, Mdust with the UV slope, βFUV. The correlation with βFUV alone accounts for 44% of this variation, which confirms that the spectral information on lines is the most important one. The other tested parameters – metallicity (about 2%) and redshift (<1%) – as well as the level of the sub-millimetre upper limits, are not significantly correlated with Mdust in this analysis.

The analysis of the results suggests that the SCUBA-2 sub-millimetre fluxes do not significantly help in constraining the dust mass (see Fig. C.3), as we do not find any correlation between the measured fluxes or upper limits. NOEMA detections are deeper and provide flux densities that are more useful in constraining the dust mass. However, we only have two objects in the sample, and none of them are within the lower sequence of GELDAs. The Balmer decrement Hα/Hβ and the dust attenuation for Hα, A(Hα) are strongly correlated with the Mdust (see Fig. 10) correlation coefficient rHα/Hβ = 0.874). The UV slope βFUV is (see Fig. 11), although at a lower level, also correlated with Mdust (correlation coefficient rβ_FUV = 0.667). We can thus conclude that, first, the emission lines, and second, the continuum shape drive the estimation of the amount of energy transferred into the far-IR. For our galaxy sample, the observed NIRSpec spectrum from about 0.5–5.3 μm provides the best spectral information to estimate the amount of dust attenuation via the Balmer decrement and the UV slope. Then, using the energy balance hypothesis, we can estimate the IR luminosity and thus Mdust, if the IR spectrum from Burgarella et al. (2022) is assumed to be valid for our present sample.

We performed tests that suggest that the minimum dust mass that we could estimate with CIGALE with the above assumptions is log10 (Mdust/M⊙) = 5.0 (Fig. 12). To perform these tests, we used CIGALE to create a mock catalogue based on the best-fit SPEDs for each object derived from a first fit. To these best-fit SPEDs, we added the observed noise drawn assuming a Gaussian distribution (see Boquien et al. 2019 for a more detailed explanation).

thumbnail Fig. 12.

Use of mock analysis to estimate minimum dust mass that we can actually estimate using our approach. Top: x-axis shows modelled Mdust computed from the best-fit models and for each of our galaxies. CIGALE is able to recover these input dust masses (y-axis) by fitting the mock data down to about log10(Mdust) = 5.0. Parts of the objects in green and all the objects in blue should thus be considered as upper limits. These objects that are the ones identified as transitioners between the upper sequence and a possible lower sequence are well below the prime sequence (red dots). Bottom: log10(Mdust) versus log10(Mstar) diagram showing that the lowest dust masses correspond to the lowest stellar masses. However, for the same stellar-mass range, we do observe a wide range of dust masses. Normal, mid, and low Mdust mean Mdust≤105 M⊙, 105 < Mdust≤106 M⊙, and Mdust > 106 M⊙, respectively.

In Fig. 13, we compare the trends related to the increase in the specific mass of metals (MZ/Mstar) and of dust (Mdust/Mstar) with cosmic ages and with the star formation rate sSFR = SFR/Mstar. Several points can be noticed: first, both follow a similar trend; second, MZ/Mstar is always higher than Mdust/Mstar (the metal mass is larger than the dust mass); Third, we observe a lack of extremely low-metallicity galaxies, as all the galaxies observed so far are above a critical metallicity value (Bayesian values derived by CIGALE) of about Zcrit = 10−3.

thumbnail Fig. 13.

Evolution of metal and dust masses. MZ/Mstar (blue dots) and Mdust/Mstar (red dots) both regularly increase with the age of the Universe (top); more metals and dust grains are formed as the Universe ages. We note that even for galaxies (blue symbols) in the early Universe (ageUniverse≲600 Myr or z ≳ 9), MZ/Mstar never goes below a few 10−3, possibly suggesting a fast rise of metals that produces the observed threshold. MZ/Mstar increases faster than Mdust/Mstar, with a much larger dispersion for Mdust/Mstar. The light blue and red areas show the mean and 2σ confidence interval of the distribution within several bins. At the bottom, same as aboce a function of sSFR. MZ/Mstar (in blue) and Mdust/Mstar (in red) decrease from high to low sSFR (bottom), as also shown in, e.g. Palla et al. (2024) and Shivaei et al. (2022). We also note a larger dispersion at lower sSFR for Mdust/Mstar, but only for Mdust/Mstar.

A quantisation of the mean dust-to-metal, DTM = 〈Mdust/MZ〉, shown in Fig. 13, gives DTMmean = 0.080, DTMmedian = 0.006 for GELDAs with a first quartile (25%) of Q1 = 0.002 and a third quartile (75%) of Q3 = 0.037; whereas for non GELDAs it is DTMmean = 0.654, DTMmedian 0.156 with Q1 = 0.085, and Q3 = 0.331. We observe a strong break in the DTM that could be interpreted (Inoue 2011; Asano et al. 2013; Zhukovska 2014; Feldmann 2015; Popping et al. 2017; Hou et al. 2019; Li et al. 2019; Graziani et al. 2020; Triani et al. 2020; Parente et al. 2022; Choban et al. 2024, and Dubois et al. 2024) as a hint that GELDAs have not started dust accretion growth, while non-GELDAs are above the critical metallicity and have dust growth in the ISM. It is interesting to note that Rémy-Ruyer et al. (2014) observed a large scatter in the gas-to-dust mass ratio for a sample of 126 galaxies spanning a 2 dex metallicity range. This scatter appears at 7.2 ≲ 12 + logOH≲8.7 and is consistent with the dust growth in the ISM predicted by Asano et al. (2013) and other works cited in this paper. The objects are those appearing on the bottom left of Figs. 6 and 13; that is, GELDAs.

4. Discussion

We now analyse the possible origins of GELDAs. To this aim, we first stacked the spectra of the GELDAs (see Fig. 14 and line fluxes extracted from the stacked spectrum in Table 2). The stacked spectrum in Fig. 14 is characteristic of star-forming galaxies, with a very blue UV slope of βFUV = –2.451 ± 0.066. This sample of GELDAs is compatible with no dust attenuation for Case B2: Hα/Hβ = 2.932 ± 0.660. To reach Hα/Hβ = 2.86 (no dust attenuation), we had to apply a correction to Hβ for the underlying absorption of 2.5%, which is a relatively low correction (Kashino et al. 2013; Reddy et al. 2015, and Shivaei et al. 2020).

thumbnail Fig. 14.

Stacked spectrum of objects selected in lower sequence. Vertical lines show the location of a few usual emission lines (see inside caption). We show the position of the lines, including where the HeI lines would be expected. None of them are detected, confirming these objects do not contain a dominant AGN. This stacked spectrum is similar to that of a young starburst, with a blue UV slope and prominent hydrogen and [OIII) emission lines.

Table 2.

Measured UV slope and fluxes of emission lines of GELDA stacked spectrum are in erg/cm2/s/Å.

The simplest origin explanation is that these GELDAs are almost unaffected by dust attenuation because they did not produce a significant dust mass. Although this might be possible at z ≳ 8.8, this is less likely at lower redshifts because of the global increase in the dust mass density that could pollute gas in the intergalactic medium (e.g. Madau & Dickinson 2014 and Traina et al. 2024) or in the average dust attenuation of galaxies (Burgarella et al. 2013 and Bogdanoska & Burgarella 2020) from the early Universe to z = 3–4. Moreover, some works (e.g. Langeroodi et al. 2024) suggest that dust is formed very fast by SNae on timescales shorter than ∼30 Myr. Even if not all dust grains had been destroyed by the SNae reverse shock, some residual attenuation of 0.05 ≤ AV ≤ 0.2 occurred, which translates to AFUV = 0.15–1.0 depending on the dust attenuation law (Salim & Narayanan 2020), and should still be detectable.

Another origin could be linked to the relative geometry of dust and stars in these objects or their small sizes. We know that the brightest H II regions in local galaxies show a correlation between the Balmer-line reddening and the dust-mass surface density (Kreckel et al. 2013; Trayford et al. 2020; Seillé et al. 2022, and Robertson et al. 2024). Our high-redshift galaxies are small (Table 2). The half-light radii measured in the F200W NIRCam images are 〈RHF200W〉 = 380±132 pc, and even smaller for z > 8.8 (〈RHF200W〉 = 327±87 pc); they are thus very dense. We measured the surface densities of gas 0≲log10(Σgas[M⊙ pc−2])≲6 and the surface densities of SFR 4≲log10(ΣSFR[M⊙ yr−1 kpc2]≲3. Star formation in galaxies is closely related to the local gas density and follows the so-called Schmidt law (Schmidt 1959). In the dense cores of star formation regions studied in the Milky Way (e.g. Shimajiri et al. 2017 and Mattern et al. 2024) and in local spiral galaxies (Gao & Solomon 2004), active high-mass star formation is intimately related to the very dense molecular gas, Mdense. However, when the density of the gas reaches an H2 surface density ΣH2≳100−200 M⊙ pc−2 (Mattern et al. 2024), a density much lower than the estimated values for our galaxies, turbulence can partially prevent star formation and reduce the SFR. This could provide us with an explanation for our sample being at a low star formation efficiency (SFE) level; that is, below other objects at the same densities in Fig. 15 and offset from both high-redshift sub-millimetre galaxies. These low SFEs are not in agreement with the suggested high-FFB-related SFE, which is one of the suggested origins of the excess of UV-bright galaxies in the early Universe.

thumbnail Fig. 15.

SFR surface density as function of gas-mass surface density for a sample of local and high-redshift galaxies, including HZ10, LBG-1, AzTEC-3, and CRLE, adapted from Pavesi et al. (2019) and Shi et al. (2011). Our objects, GELDAs, and non-GELDAs are very dense, and all of them are found slightly below other objects, which might be related to the very high gas density in these objects where the turbulence can produce a negative effect on the SFE, which is thus lower than in other high-redshift galaxies at the same gas surface density. From bottom to top, the shaded areas correspond to low-surface-brightess (magenta), late-type (blue), early type (green), hi-z star-forming (yellow), and hi-z sub-millimetre (red) galaxies. They are also given inside the figure.

A third possible explanation can be found in Ferrara et al. (2023), where a model was built that reproduces the excess of UV-bright galaxies in the luminosity functions at z = 10–14. The authors proposed that these galaxies at z ≳ 11 contain negligible amounts of dust and that most of the dust produced by SNae in these objects could have been efficiently ejected during the very first phases of galaxy build-up; that is because these galaxies are bursty, and they are able to temporarily evacuate large amounts of gas and dust far from the star-forming region (e.g. Sun et al. 2023 and Choban et al. 2024). In this case, the lower objects could correspond to galaxies where dust is efficiently ejected far from the stellar populations by radiation pressure as soon as it is produced by stars. However, if winds had ejected dust in high-redshift galaxies, they probably also expelled gas, and for both GELDAs and non-GELDAs, we find fgas = Mgas/(Mgas+Mstar) = 0.96 ± 0.03, in agreement with models predicting that galaxies with log10(Mstar/M⊙)<9.0 have fgas≳0.75 (Davé et al. 2017 and Popping et al. 2014). Yet, galaxies at z > 8.8 have a slightly lower but still very high value of fgas = 0.90±0.05. Thus, these gas fractions show that these objects still contain a large mass of gas and should therefore also contain dust.

Finally, in the Mdust versus Mstar diagram plotted in Fig. 16, we observe a concentration of galaxies in the upper part of the figure, mainly in the range of 8.0<log10(Mstar)<10.0. We also see a significant decline in dust attenuation at log10(Mstar) ∼ 8.0–9.0 that was already seen in Fig. 6. This effect means that Mdust is significantly lower by a factor of 100–1000 at a given stellar mass, with a lower clump or sequence well below the upper one. This biphasic plot could suggest a two-mode building of dust mass in galaxies.

thumbnail Fig. 16.

Mdust as function of Mstar. Top: Dots are colour-coded according to redshift. The size of the symbols provides us with information on 12 + log10(O/H), with the largest sizes corresponding to the largest metallicities. We show density contours as heavy black lines. At log10 (Mstar) ∼ 108−9 M⊙, we observe a transition from an apparent high sequence to a lower one. The upper sequence is similar to that observed at low redshift (e.g. Beeston et al. 2018). The lower objects have not been identified before in this (Mdust vs. Mstar) diagram, except for a few objects (green crosses, De Vis et al. 2017). They share the same location in the plot and are extracted from a sub-sample of galaxies in an HI-selected local-galaxy sample from GAMA and H-ATLAS. Bottom: Models plotted on density contours. Hydrodynamical code dustygadget (Graziani et al. 2020) is in yellow, a cosmological hydrodynamical simulation coupled with a chemical evolution model (Mancini et al. 2015) is in green, and a suite of cosmological, fluid-dynamical simulations of galaxies (Esmerian & Gnedin 2024) are in pink. The upper brown line (Witstok et al. 2023) shows where galaxies with ISM-grown grains should be. The bottom orange line corresponds to galaxies with only stardust (Witstok et al. 2023) that underwent a 95% destruction of grains by reverse SNae shock (Witstok et al. 2023). The objects observed at the bottom of the top panel correspond to the transition between galaxies only containing stardust to galaxies dominated by ISM dust.

To demonstrate the nature of this lower sequence, Fig. 16 shows several models (Mancini et al. 2015; Graziani et al. 2020; Esmerian & Gnedin 2024, and Witstok et al. 2023). The first dust grains should have formed in stellar ejecta from SNae (and maybe AGB stars). However, a possibly substantial fraction of these dust grains is likely destroyed by the SNae reverse shock. After this first phase, the remaining dust grains form seeds and accrete ISM material for grain growth. This process seems to happen only when a critical ISM metallicity is reached at 0.05≲Z/Z⊙≲0.5 (Inoue 2011; Asano et al. 2013; Zhukovska 2014; Feldmann 2015; Popping et al. 2017; Hou et al. 2019; Li et al. 2019; Graziani et al. 2020; Triani et al. 2020; Parente et al. 2022, and Choban et al. 2024). While the upper sequence would have a dust mass where grains have grown in the ISM, the lower sequence would correspond to stardust grains only formed from SNae, with a grain destruction rate by the SNae reverse shock of the order of 95% (Witstok et al. 2023). The jump from the lower to the upper sequence predicted by the models agrees well with our data. If this population of GELDAs only contains stardust, that would provide a natural explanation for the excess of UV-bright galaxies at z < 10 detected by the JWST. We check in Figs. C.1 and C.3 that other assumptions lead to the same apparent transition. Only when no spectroscopic data is used in the fits does the shape of the transition change. Finally, Fig. 17 shows that there are significantly fewer metals in GELDAs compared to the rest of the sample. In this plot, the GELDAs are found in the bottom left of the figure with a metal mass log10(MZ) ≲ 7.5, while the sample range extends beyond 5.5 ≲ log10(MZ) ≲ 9.0. This is expected if these objects did not undergo any growth of the dust grains triggered by a larger amount of metal in the ISM, because this accretion of ISM material for grain growth is triggered when the metallicity reaches a minimum critical threshold (Asano et al. 2013).

thumbnail Fig. 17.

Objects colour-coded by redshift. AGNs are shown as crosses, and GELDAs are upside-down Ys. This figure shows that GELDAs (both z ⩾ 8.8 in red and z < 8.8 in blue) have lower metal masses than the rest of the sample. This would support the hypothesis that the origin of dust grains in these objects might not be due to ISM growth.

To test whether our hypothesis could be correct, we tried to check if we could observe a difference in metallicity for GELDAs and non-GELDAs in Fig. 18. The Kolmogorov-Smirnov (KS) test shown in Fig. 18 suggests that the difference between the two distributions (GELDAs and non-GELDAS) is highly significant. Furthermore, a metallicity threshold is found at 12 + log10(O/H) = 7.60, which corresponds (with Z⊙ = 0.014 and Eq. (1)) to Z = 0.11. This is in excellent agreement with the critical metallicity (i.e. the metallicity at which the contribution of stars equals that of the dust-mass growth in the ISM) predicted by the models shown in Fig. 18 and with most models listed in this paper. For example, Asano et al. (2013) gave Z/Z⊙ = 0.2; that is, 12 + log10(O/H) = 7.86, and Feldmann et al. (2025) give Z/Z⊙ = 0.1; that is, 12 + log10(O/H) = 7.56.

thumbnail Fig. 18.

Distribution of metallicity for GELDAs and non-GELDAs. The left panel presents the kernel density estimation (KDE) with an Epanechnikov kernel and a bandwidth = 0.5 both for GELDAs (solid blue line) and non-GELDAS (solid red line). The dashed black lines show the differences. The blue and red shaded areas show where GELDAs and non-GELDAs are the dominant population, respectively. We clearly see that GELDAs preferentially cluster at low metallicity, with a threshold estimated at 12 + log10(O/H) = 7.60. The Kolmogorov-Smirnov test confirms that the difference is highly significant. We show predictions for the transition from stardust to ISM from some of the models (see main text for more works): Asano et al. (2013) (orange shaded area), Zhukovska (2014) (green shaded area), and Feldmann (2015) (dashed orange line). Rémy-Ruyer et al. (2014) found a larger scatter in their dust-to-gas ratio covers almost the same metallicity range: 7.2 ≲ 12 + log10(O/H) ≲ 8.5. All of these metallicities are in good agreement with our estimated threshold. The middle and right panels show the rolling averages (window size = 15) of the metallicity as a function of AFUV and log10 (Mdust). Both panels confirm the difference between GELDAs (blue) and non-GELDAS (red).

The present results provide us with an inherent explanation for the UV-bright tension in the early Universe if these galaxies contain only a low dust mass mainly formed in the circumstellar medium around SNae in the very first phases of star formation, and not by accretion in the ISM. This hypothesis is further supported by the fact that in our total sample, 5/6 galaxies, that is, ∼83.3% of the z ⩾ 8.8 galaxies, are GELDAs; whereas, at z < 8.8 only 38/167, that is, 22.8%, are GELDAs, suggesting this type of galaxy could become dominant in the early Universe.

5. Conclusions

We detected a population of galaxies with extremely low dust attenuation (GELDAs) in the 4.0 < z < 11.4 redshift range using a new version of the CIGALE code that accepts both photometric and spectroscopic data. GELDAs are defined as follows:

  • AFUV = 0 within 2σA_FUV, that is no dust attenuation

  • Mstar < 109 M⊙.

The present galaxy sample shares most of its properties with the ALPINE one in terms of stellar mass (8.0 ≲ Mstar≲11.4) and redshift (two redshift ranges at 4.40 < z < 4.65 and 5.05 < z < 5.90, Burgarella et al. 2022). Assuming that the far-IR dust emission of our sample is similar to that of the ALPINE galaxy sample, we estimated the dust masses. In the Mdust versus Mstar diagram, we clearly see a transition at log10(Mstar) ∼ 8.5 between an upper and a lower sequence. A comparison with models suggests that the transition galaxies could mark the shift from dust solely produced by stellar evolution (stardust galaxies) to dust growth in the ISM of galaxies. The dust-to-metal ratios are very low for GELDAs (DTM mean GELDA = 0.080 $ {}_{\mathrm {mean}}^{\mathrm {GELDA}} = 0.080 $, DTM median GELDA = 0.006 ) $ {}_{\mathrm {median}}^{\mathrm {GELDA}} = 0.006) $, with a third quartile Q3 = 0.037, and quite high (DTM mean non − GELDA = 0.654 $ {}_{\mathrm {mean}}^{\mathrm {non-GELDA}} = 0.654 $, DTM median non − GELDA = 0.156 ) , $ {}_{\mathrm {median}}^{\mathrm {non-GELDA}} = 0.156), $ with Q3 = 0.331 for a non-GELDA; this is in agreement with the hypothesis of stardust versus ISM dust.

In our data, a KS test suggests this transition appears at 12 + log10(O/H) = 7.60 (Z/Z⊙ = 0.1), which is in excellent agreement with the predicted metallicity at which the contribution of stars would equal that of the growth of the dust mass in the ISM in Asano et al. (2013). The gas-mass fraction is fgas=Mgas/(Mgas+Mstar) > 0.9 for our entire sample of galaxies, including GELDAs at all redshifts. This suggests that there is a large gas mass in the galaxies that was not expelled, and it supports the hypothesis that dust formed in the galaxies should still remain inside them.

Finally, the SFE of our galaxies is in agreement with the Schmidt-Kennicutt law (Kennicutt 1998), although at lower SFEs than high-redshift sub-millimetre galaxies at the same gas-mass density. In the highest redshift bin at z > 8.8, almost all galaxies (∼83.3%) can be assigned to the GELDA category, while less than 1/4 of the low-redshift (z < 8.8) galaxies are GELDAs. These different regimes might mark a transition around z ∼ 9. In this highest redshift bin, galaxies with low Mdust/Mstar and blue UV slopes contain young, metal-poor stars that may be forming their first dust grains from pop. II and at z>9, possibly pop. III stars, along with their first metals.

Such stardust galaxies are ideal for explaining the excess production of UV-bright galaxies in the early Universe, because they might become dominant in the early Universe. We do not detect any extremely low-metallicity values above zgas∼10−3 in our sample of galaxies, even at z ≳ 8.8, suggesting either a bias in our sample or a rapid rise of metals in the early Universe.

Acknowledgments

DB and VB thank the Programme National Cosmology and Galaxies (PNCG) and the Centre National d’Etudes Spatiales (CNES) for their financial support.

1

The Cosmic Evolution Early Release Science Survey; https://ceers.github.io/.

2

From Groves et al. (2012): Case A and Case B. Case A assumes that an ionised nebula is optically thin to all Lyman emission lines, while Case B assumes that a nebula is optically thick to all Lyman lines greater than Lyα, meaning these photons are absorbed and re-emitted as a combination of Lyα and higher order lines, such as the Balmer lines. These two cases will lead to different intrinsic ratios for the Balmer lines, with variations of the same order as temperature effects. Although Case B is typically assumed for determining intrinsic ratios, in reality the ratio in typical H II regions lies between these two cases.

References

  1. Akins, H. B., Casey, C. M., Allen, N., et al. 2023, ApJ, 956, 61 [NASA ADS] [CrossRef] [Google Scholar]
  2. Algera, H. S. B., Inami, H., Sommovigo, L., et al. 2024a, MNRAS, 527, 6867 [Google Scholar]
  3. Algera, H. S. B., Inami, H., De Looze, I., et al. 2024b, MNRAS, 533, 3098 [NASA ADS] [CrossRef] [Google Scholar]
  4. Arrabal Haro, P., Dickinson, M., Finkelstein, S. L., et al. 2023, ApJ, 951, L22 [NASA ADS] [CrossRef] [Google Scholar]
  5. Asano, R. S., Takeuchi, T. T., Hirashita, H., & Inoue, A. K. 2013, Earth Planets Space, 65, 213 [NASA ADS] [CrossRef] [Google Scholar]
  6. Atek, H., Shuntov, M., Furtak, L. J., et al. 2023, MNRAS, 519, 1201 [Google Scholar]
  7. Bagley, M. B., Finkelstein, S. L., Koekemoer, A. M., et al. 2023, ApJ, 946, L12 [NASA ADS] [CrossRef] [Google Scholar]
  8. Beeston, R. A., Wright, A. H., Maddox, S., et al. 2018, MNRAS, 479, 1077 [NASA ADS] [Google Scholar]
  9. Béthermin, M., Fudamoto, Y., Ginolfi, M., et al. 2020, A&A, 643, A2 [Google Scholar]
  10. Birkmann, S. M., Ferruit, P., Giardino, G., et al. 2022, A&A, 661, A83 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  11. Bogdanoska, J., & Burgarella, D. 2020, MNRAS, 496, 5341 [NASA ADS] [CrossRef] [Google Scholar]
  12. Boquien, M., Burgarella, D., Roehlly, Y., et al. 2019, A&A, 622, A103 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  13. Bouwens, R. J., Aravena, M., Decarli, R., et al. 2016, ApJ, 833, 72 [NASA ADS] [CrossRef] [Google Scholar]
  14. Bruzual, G., & Charlot, S. 2003, MNRAS, 344, 1000 [NASA ADS] [CrossRef] [Google Scholar]
  15. Burgarella, D., Buat, V., & Iglesias-Páramo, J. 2005, MNRAS, 360, 1413 [NASA ADS] [CrossRef] [Google Scholar]
  16. Burgarella, D., Buat, V., Gruppioni, C., et al. 2013, A&A, 554, A70 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  17. Burgarella, D., Bogdanoska, J., Nanni, A., et al. 2022, A&A, 664, A73 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  18. Burgarella, D., Buat, V., Boquien, M., Roehlly, Y., & Lambert, J. C. 2024, in Observatory Operations: Strategies, Processes, and Systems X, eds. C. R. Benn, A. Chrysostomou, & L. J. Storrie-Lombardi, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, 13098, 130980N [Google Scholar]
  19. Casey, C. M., Akins, H. B., Shuntov, M., et al. 2024, ApJ, 965, 98 [NASA ADS] [CrossRef] [Google Scholar]
  20. Chabrier, G. 2003, PASP, 115, 763 [Google Scholar]
  21. Chakraborty, P., Sarkar, A., Wolk, S., et al. 2024, MNRAS, submitted [arXiv e-prints, arXiv:2406.05306] [Google Scholar]
  22. Choban, C. R., Kereš, D., Sandstrom, K. M., et al. 2024, MNRAS, 529, 2356 [CrossRef] [Google Scholar]
  23. Chowdhury, A., Kanekar, N., & Chengalur, J. N. 2022, ApJ, 935, L5 [NASA ADS] [CrossRef] [Google Scholar]
  24. Cole, J. W., Papovich, C., Finkelstein, S. L., et al. 2025, ApJ, 979, 193 [Google Scholar]
  25. Davé, R., Rafieferantsoa, M. H., Thompson, R. J., & Hopkins, P. F. 2017, MNRAS, 467, 115 [NASA ADS] [Google Scholar]
  26. Dayal, P., Ferrara, A., Sommovigo, L., et al. 2022, MNRAS, 512, 989 [NASA ADS] [CrossRef] [Google Scholar]
  27. De Vis, P., Dunne, L., Maddox, S., et al. 2017, MNRAS, 464, 4680 [NASA ADS] [CrossRef] [Google Scholar]
  28. De Vis, P., Jones, A., Viaene, S., et al. 2019, A&A, 623, A5 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  29. Dell’Agli, F., Valiante, R., Kamath, D., Ventura, P., & García-Hernández, D. A. 2019, MNRAS, 486, 4738 [Google Scholar]
  30. Di Cesare, C., Graziani, L., Schneider, R., et al. 2023, MNRAS, 519, 4632 [Google Scholar]
  31. Draine, B. T., Aniano, G., Krause, O., et al. 2014, ApJ, 780, 172 [Google Scholar]
  32. Dubois, Y., Rodríguez Montero, F., Guerra, C., et al. 2024, A&A, 687, A240 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  33. Esmerian, C. J., & Gnedin, N. Y. 2024, ApJ, 968, 113 [NASA ADS] [CrossRef] [Google Scholar]
  34. Euclid Collaboration (Bisigello, L., et al.) 2023, MNRAS, 520, 3529 [NASA ADS] [CrossRef] [Google Scholar]
  35. Feldmann, R. 2015, MNRAS, 449, 3274 [NASA ADS] [CrossRef] [Google Scholar]
  36. Feldmann, R., Boylan-Kolchin, M., Bullock, J. S., et al. 2025, MNRAS, 536, 988 [Google Scholar]
  37. Fernández, V., Amorín, R., Firpo, V., & Morisset, C. 2024, A&A, 688, A69 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  38. Ferrara, A., Sommovigo, L., Dayal, P., et al. 2022, MNRAS, 512, 58 [NASA ADS] [CrossRef] [Google Scholar]
  39. Ferrara, A., Pallottini, A., & Dayal, P. 2023, MNRAS, 522, 3986 [NASA ADS] [CrossRef] [Google Scholar]
  40. Ferrara, A., Pallottini, A., & Sommovigo, L. 2025, A&A, 694, A286 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  41. Finkelstein, S. L., Bagley, M. B., Ferguson, H. C., et al. 2023, ApJ, 946, L13 [NASA ADS] [CrossRef] [Google Scholar]
  42. Finkelstein, S. L., Leung, G. C. K., Bagley, M. B., et al. 2024, ApJ, 969, L2 [NASA ADS] [CrossRef] [Google Scholar]
  43. Fudamoto, Y., Ivison, R. J., Oteo, I., et al. 2017, MNRAS, 472, 2028 [NASA ADS] [CrossRef] [Google Scholar]
  44. Gao, Y., & Solomon, P. M. 2004, ApJ, 606, 271 [NASA ADS] [CrossRef] [Google Scholar]
  45. Graziani, L., Schneider, R., Ginolfi, M., et al. 2020, MNRAS, 494, 1071 [NASA ADS] [CrossRef] [Google Scholar]
  46. Grieco, F., Theulé, P., De Looze, I., & Dulieu, F. 2023, Nat. Astron., 7, 541 [NASA ADS] [CrossRef] [Google Scholar]
  47. Grogin, N. A., Kocevski, D. D., Faber, S. M., et al. 2011, ApJS, 197, 35 [NASA ADS] [CrossRef] [Google Scholar]
  48. Groves, B., Brinchmann, J., & Walcher, C. J. 2012, MNRAS, 419, 1402 [NASA ADS] [CrossRef] [Google Scholar]
  49. Harikane, Y., Zhang, Y., Nakajima, K., et al. 2023, ApJ, 959, 39 [NASA ADS] [CrossRef] [Google Scholar]
  50. Heintz, K. E., Shapley, A. E., Sanders, R. L., et al. 2023a, A&A, 678, A30 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  51. Heintz, K. E., Giménez-Arteaga, C., Fujimoto, S., et al. 2023b, ApJ, 944, L30 [NASA ADS] [CrossRef] [Google Scholar]
  52. Hezaveh, Y. D., Marshall, P. J., & Blandford, R. D. 2015, ApJ, 799, L22 [Google Scholar]
  53. Hirashita, H., Ferrara, A., Dayal, P., & Ouchi, M. 2014, MNRAS, 443, 1704 [NASA ADS] [CrossRef] [Google Scholar]
  54. Hopkins, P. F., Wetzel, A., Kereš, D., et al. 2018, MNRAS, 480, 800 [NASA ADS] [CrossRef] [Google Scholar]
  55. Hou, K. -C., Aoyama, S., Hirashita, H., Nagamine, K., & Shimizu, I. 2019, MNRAS, 485, 1727 [NASA ADS] [CrossRef] [Google Scholar]
  56. Hutter, A., Cueto, E. R., Dayal, P., et al. 2025, A&A, 694, A254 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  57. Ilie, C., Paulin, J., & Freese, K. 2023, Proc. Natl. Acad. Sci., 120, e2305762120 [Google Scholar]
  58. Inami, H., Algera, H. S. B., Schouws, S., et al. 2022, MNRAS, 515, 3126 [NASA ADS] [CrossRef] [Google Scholar]
  59. Inoue, A. K. 2011, Earth Planets Space, 63, 1027 [NASA ADS] [CrossRef] [Google Scholar]
  60. Iyer, K. G., Gawiser, E., Faber, S. M., et al. 2019, ApJ, 879, 116 [NASA ADS] [CrossRef] [Google Scholar]
  61. Jeong, T. B., Jeon, M., Song, H., & Bromm, V. 2025, ApJ, 980, 10 [NASA ADS] [CrossRef] [Google Scholar]
  62. Kashino, D., Silverman, J. D., Rodighiero, G., et al. 2013, ApJ, 777, L8 [Google Scholar]
  63. Kennicutt, R. C. Jr. 1998, ApJ, 498, 541 [Google Scholar]
  64. Kocevski, D. D., Onoue, M., Inayoshi, K., et al. 2023, ApJ, 954, L4 [NASA ADS] [CrossRef] [Google Scholar]
  65. Koekemoer, A. M., Faber, S. M., Ferguson, H. C., et al. 2011, ApJS, 197, 36 [NASA ADS] [CrossRef] [Google Scholar]
  66. Komatsu, E., Smith, K. M., Dunkley, J., et al. 2011, ApJS, 192, 18 [Google Scholar]
  67. Kreckel, K., Groves, B., Schinnerer, E., et al. 2013, ApJ, 771, 62 [NASA ADS] [CrossRef] [Google Scholar]
  68. Langeroodi, D., Hjorth, J., Ferrara, A., & Gall, C. 2024, arXiv e-prints [arXiv:2410.14671] [Google Scholar]
  69. Laporte, N., Ellis, R. S., Boone, F., et al. 2017, ApJ, 837, L21 [CrossRef] [Google Scholar]
  70. Larson, R. L., Finkelstein, S. L., Kocevski, D. D., et al. 2023, ApJ, 953, L29 [NASA ADS] [CrossRef] [Google Scholar]
  71. Lei, L., Wang, Y. -Y., Yuan, G. -W., et al. 2025, ApJ, 980, 249 [Google Scholar]
  72. Leja, J., Carnall, A. C., Johnson, B. D., Conroy, C., & Speagle, J. S. 2019, ApJ, 876, 3 [Google Scholar]
  73. Leśniewska, A., & Michałowski, M. J. 2019, A&A, 624, L13 [Google Scholar]
  74. Lewis, J. S. W., Ocvirk, P., Dubois, Y., et al. 2023, MNRAS, 519, 5987 [CrossRef] [Google Scholar]
  75. Li, Q., Narayanan, D., & Davé, R. 2019, MNRAS, 490, 1425 [CrossRef] [Google Scholar]
  76. Li, Z., Dekel, A., Sarkar, K. C., et al. 2024, A&A, 690, A108 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  77. Lower, S., Narayanan, D., Leja, J., et al. 2020, ApJ, 904, 33 [NASA ADS] [CrossRef] [Google Scholar]
  78. Madau, P., & Dickinson, M. 2014, ARA&A, 52, 415 [Google Scholar]
  79. Małek, K., Buat, V., Roehlly, Y., et al. 2018, A&A, 620, A50 [Google Scholar]
  80. Mancini, M., Schneider, R., Graziani, L., et al. 2015, MNRAS, 451, L70 [CrossRef] [Google Scholar]
  81. Mattern, M., André, P., Zavagno, A., et al. 2024, A&A, 688, A163 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  82. Mauerhofer, V., & Dayal, P. 2023, MNRAS, 526, 2196 [NASA ADS] [CrossRef] [Google Scholar]
  83. Messias, H., Guerrero, A., Nagar, N., et al. 2024, MNRAS, 533, 3937 [Google Scholar]
  84. Naidu, R. P., Oesch, P. A., van Dokkum, P., et al. 2022, ApJ, 940, L14 [NASA ADS] [CrossRef] [Google Scholar]
  85. Nakajima, K., Ouchi, M., Xu, Y., et al. 2022, ApJS, 262, 3 [CrossRef] [Google Scholar]
  86. Nakajima, K., Ouchi, M., Isobe, Y., et al. 2023, ApJS, 269, 33 [NASA ADS] [CrossRef] [Google Scholar]
  87. Nanni, A., Burgarella, D., Theulé, P., Côté, B., & Hirashita, H. 2020, A&A, 641, A168 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  88. Nicholls, C. P., Lebzelter, T., Smette, A., et al. 2017, A&A, 598, A79 [EDP Sciences] [Google Scholar]
  89. Palla, M., De Looze, I., Relaño, M., et al. 2024, MNRAS, 528, 2407 [NASA ADS] [CrossRef] [Google Scholar]
  90. Parente, M., Ragone-Figueroa, C., Granato, G. L., et al. 2022, MNRAS, 515, 2053 [NASA ADS] [CrossRef] [Google Scholar]
  91. Pavesi, R., Riechers, D. A., Faisst, A. L., Stacey, G. J., & Capak, P. L. 2019, ApJ, 882, 168 [Google Scholar]
  92. Popping, G., Somerville, R. S., & Trager, S. C. 2014, MNRAS, 442, 2398 [NASA ADS] [CrossRef] [Google Scholar]
  93. Popping, G., Somerville, R. S., & Galametz, M. 2017, MNRAS, 471, 3152 [NASA ADS] [CrossRef] [Google Scholar]
  94. Pozzi, F., Calura, F., Fudamoto, Y., et al. 2021, A&A, 653, A84 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  95. Reddy, N. A., Kriek, M., Shapley, A. E., et al. 2015, ApJ, 806, 259 [NASA ADS] [CrossRef] [Google Scholar]
  96. Rémy-Ruyer, A., Madden, S. C., Galliano, F., et al. 2014, A&A, 563, A31 [Google Scholar]
  97. Robertson, C., Holwerda, B. W., Young, J., et al. 2024, AJ, 167, 263 [CrossRef] [Google Scholar]
  98. Salim, S., & Narayanan, D. 2020, ARA&A, 58, 529 [NASA ADS] [CrossRef] [Google Scholar]
  99. Sanders, R. L., Shapley, A. E., Topping, M. W., Reddy, N. A., & Brammer, G. B. 2024, ApJ, 962, 24 [NASA ADS] [CrossRef] [Google Scholar]
  100. Schmidt, M. 1959, ApJ, 129, 243 [NASA ADS] [CrossRef] [Google Scholar]
  101. Schneider, R., & Maiolino, R., 2024, A&A Rev., 32, 2 [Google Scholar]
  102. Seillé, L. M., Buat, V., Haddad, W., et al. 2022, A&A, 665, A137 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  103. Shi, Y., Helou, G., Yan, L., et al. 2011, ApJ, 733, 87 [NASA ADS] [CrossRef] [Google Scholar]
  104. Shimajiri, Y., André, P., Braine, J., et al. 2017, A&A, 604, A74 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  105. Shivaei, I., Reddy, N., Rieke, G., et al. 2020, ApJ, 899, 117 [NASA ADS] [CrossRef] [Google Scholar]
  106. Shivaei, I., Popping, G., Rieke, G., et al. 2022, ApJ, 928, 68 [NASA ADS] [CrossRef] [Google Scholar]
  107. Sommovigo, L., Ferrara, A., Carniani, S., et al. 2022a, MNRAS, 517, 5930 [NASA ADS] [CrossRef] [Google Scholar]
  108. Sommovigo, L., Ferrara, A., Pallottini, A., et al. 2022b, MNRAS, 513, 3122 [NASA ADS] [CrossRef] [Google Scholar]
  109. Speagle, J. S., Steinhardt, C. L., Capak, P. L., & Silverman, J. D. 2014, ApJS, 214, 15 [Google Scholar]
  110. Strandet, M. L., Weiss, A., De Breuck, C., et al. 2017, ApJ, 842, L15 [Google Scholar]
  111. Sun, G., Faucher-Giguère, C. -A., Hayward, C. C., & Shen, X. 2023, MNRAS, 526, 2665 [NASA ADS] [CrossRef] [Google Scholar]
  112. Tacchella, S., Johnson, B. D., Robertson, B. E., et al. 2023, MNRAS, 522, 6236 [NASA ADS] [CrossRef] [Google Scholar]
  113. Tacconi, L. J., Genzel, R., & Sternberg, A. 2020, ARA&A, 58, 157 [NASA ADS] [CrossRef] [Google Scholar]
  114. Tamura, Y., Mawatari, K., Hashimoto, T., et al. 2019, ApJ, 874, 27 [NASA ADS] [CrossRef] [Google Scholar]
  115. Theulé, P., Burgarella, D., Buat, V., et al. 2024, A&A, 682, A119 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  116. Traina, A., Magnelli, B., Gruppioni, C., et al. 2024, A&A, 690, A84 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  117. Trayford, J. W., Lagos, C. d. P., Robotham, A. S. G., & Obreschkow, D. 2020, MNRAS, 491, 3937 [NASA ADS] [CrossRef] [Google Scholar]
  118. Triani, D. P., Sinha, M., Croton, D. J., Pacifici, C., & Dwek, E. 2020, MNRAS, 493, 2490 [NASA ADS] [CrossRef] [Google Scholar]
  119. Valentino, F., Fujimoto, S., Giménez-Arteaga, C., et al. 2024, A&A, 685, A138 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  120. Valiante, R., Schneider, R., Bianchi, S., & Andersen, A. C. 2009, MNRAS, 397, 1661 [CrossRef] [Google Scholar]
  121. Vieira, J. D., Marrone, D. P., Chapman, S. C., et al. 2013, Nature, 495, 344 [Google Scholar]
  122. Vijayan, A. P., Clay, S. J., Thomas, P. A., et al. 2019, MNRAS, 489, 4072 [NASA ADS] [CrossRef] [Google Scholar]
  123. Walter, F., Carilli, C., Neeleman, M., et al. 2020, ApJ, 902, 111 [Google Scholar]
  124. Watson, D., Christensen, L., Knudsen, K. K., et al. 2015, Nature, 519, 327 [Google Scholar]
  125. Weibel, A., Oesch, P. A., Barrufet, L., et al. 2024, MNRAS, 533, 1808 [NASA ADS] [CrossRef] [Google Scholar]
  126. Witstok, J., Jones, G. C., Maiolino, R., Smit, R., & Schneider, R. 2023, MNRAS, 523, 3119 [NASA ADS] [CrossRef] [Google Scholar]
  127. Yung, L. Y. A., Somerville, R. S., Finkelstein, S. L., Wilkins, S. M., & Gardner, J. P. 2024, MNRAS, 527, 5929 [Google Scholar]
  128. Zackrisson, E., Rydberg, C. -E., Schaerer, D., Östlin, G., & Tuli, M. 2011, ApJ, 740, 13 [NASA ADS] [CrossRef] [Google Scholar]
  129. Zavala, J. A., Aretxaga, I., Dunlop, J. S., et al. 2018, MNRAS, 475, 5585 [Google Scholar]
  130. Zavala, J. A., Buat, V., Casey, C. M., et al. 2023, ApJ, 943, L9 [NASA ADS] [CrossRef] [Google Scholar]
  131. Zhukovska, S. 2014, A&A, 562, A76 [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
  132. Zhukovska, S., Henning, T., & Dobbs, C. 2018, ApJ, 857, 94 [Google Scholar]

Appendix A: Description of the spectro-photometric CIGALE

The concept of CIGALE was developed in the original paper (Burgarella et al. 2005) where multi-wavelength data from the far-UV to the far-IR could be used to derive physical parameters by fitting photometric SEDs. The present open and public version 2025.0, 17 January 2025 of CIGALE (Boquien et al. 2019) is written in Python and parallelized. It also includes a much larger number of modules (that is, physical processes and models) that provide the user with a rich choice to adapt the modeling phase to most galaxies. This Python CIGALE code is also one of the fastest SED fitting codes in the world (Burgarella et al. 2024), making it faster than some of the machine-learning-based codes (namely the convolutional neural network and the deep learning neural network as used in Euclid Collaboration: Bisigello et al. 2023). However, the most important difference of this new version is the possibility to combine spectroscopic data to photometric data (hereafter spectrophotometric data or SPED) with their own uncertainties, while conserving CIGALE's ability to fit several thousands of objects in a reasonable time. This means that whatever the parameters derived via the fitting process are, these parameters have to be consistent with both photometric and spectroscopic data. Fitting the 173 galaxies using 800 million models from this sample takes about 12 hours on a 48-core computer with 512 GB of memory, utilizing about 60 GB for each run.

In order to combine the two above data types, we have to normalize the spectrum to the photometry. We provide three options: 1) no normalization: raw data are combined, 2) we integrate the modeled spectra into the filters and estimate a global normalization factor through a χ2 when the signal-to-noise ratio >5 for the photometric bands used to compute χ2, and 3) we determine a wavelength-dependent normalization. We stress that normalizing the spectroscopic data to the photometric data could be problematic if the emission inside the photometric aperture is physically different from the emission inside the spectral slit. In this case, the resulting fit might not be realistic because of the different natures of the emitting regions. For instance, a dusty galaxy might present a clear region in the outskirts that could dominate the spectrum in UV but not elsewhere in the spectrum if both observations are not at the same position. Converging would thus be difficult, and a good global fit would not be reached. We therefore recommend being careful when combining the spectroscopic and photometric data. For small galaxies like ours, this issue is minimized because we are more likely to observe the same region photometrically and spectroscopically.

In order to simultaneously fit all the data, we need to inform CIGALE about the (sometimes wavelength-dependent) spectral resolution of the spectrometer to create resolution elements corresponding to the instrumental spectral resolution where the models are integrated. This phase is transparent to the user and is performed during the configuration of the CIGALE environment. A typical CIGALE spectrophotometric run appears similar to a photometric run from the user's point of view.

CIGALE learns that some spectroscopic data have to be taken into account from the configuration file, pcigale.ini, where a specific flag is set to ’True’ as shown in Fig. A.1.

thumbnail Fig. A.1.

This flag must be set to ”True” in the pcigale.ini file to fit spectroscopic data.

Moreover, the input table must contain the following information:

  • ”id” that contains an alphanumeric identifier (a different one for each object to be fitted)

  • ”redshift” that contains the redshift of the objects or ”NaN” if redshifts have to be estimated.

  • ”spectrum” which contains the path to the spectrum

  • ”mode” that contains the type of spectrum, e.g. ”prism” if JWST-NIRSpec data is used

  • ”norm” where one of the three normalizations should be provided for each object: ”none”, ”global” or ”wave”. Note that the normalization could be different for each object. If no photometric data are given to CIGALE, ”none” should be used, and only the spectrum will be fitted.

Appendix B: Parameters used in CIGALE's final fit

All the spectral models computed by this new version of CIGALE (momentarily dubbed CIGALE-SPEC) using the selected modules (each one corresponding to a physical emission) are convolved with NIRSpec's prism spectral resolution matched to the observed spectra. The modules and the list of priors are listed in Tab. B.1. These spectral data are added to the photometry to form a SPED. The rest of the process follows the usual flow of CIGALE as described in Boquien et al. (2019). The nebular models have been computed with CLOUDY, as described in Theulé et al. (2024). In this analysis, we normalize the prism spectroscopic data by computing a wavelength-dependent normalization, which is estimated from the photometry in the wavelength range, for the photometric bands that have a SNR>5.0. A one-dimensional piecewise linear interpolation with given discrete photometric data points is used to derive the wavelength-dependent normalization factor. This normalization is computed for each and every object with photometric data and applied to each spectroscopic observation. In this work, we used the WMAP7 cosmology (Komatsu et al. 2011). We assumed a Chabrier initial mass function (IMF, Chabrier 2003) with lower and upper mass cutoffs Mlow = 0.1 M⊙ and Mup = 100 M⊙, and a solar metallicity Z⊙ = 0.014 (Bruzual & Charlot 2003), and the dust emission models, use κν = 0.637 m2 kg−2. We show examples of SPED fits in Figs. B.1, B.3 and B.5 over the full spectral range and B.2, B.4 and B.6 over the NIRSpec range.

thumbnail Fig. B.1.

a) Objects in the upper sequence (larger Mdust) in Mdust vs. Mstar plot. We present a sample of spectral fits over the whole spectral range, that is including NIRSpec spectroscopy and the sub-millimeter data (mostly upper limits).

thumbnail Fig. B.2.

b) Same as Fig. B.1 for the same objects but only the fits of the NIRSpec spectrum is shown, which is a zoom in the previous plots.

thumbnail Fig. B.3.

Same as previous Fig. B.1.

thumbnail Fig. B.4.

Same as Fig. B.2.

thumbnail Fig. B.5.

Objects in the lower sequence (lower Mdust) in Mdust vs. Mstar plot. Same as Fig. B.1 but for objects in the lower sequence in the Mdust vs. Mstar plot.

thumbnail Fig. B.6.

Same as Fig. B.5 for the same objects but only the fits of the NIRSpec spectrum is shown, which is a zoom in the previous plots.

Table B.1.

CIGALE modules and input parameters used for all the fits. BC03 means Bruzual & Charlot (2003), and the Chabrier IMF refers to Chabrier (2003).

Appendix C: Results assuming a periodic star formation history

We present the same analysis obtained in the main paper, but here, we assume a periodic SFH, that is, a series of regular bursts over the age of galaxies (Fig. C.2). The main parameters that define the SFH for this CIGALE run are listed in Tab. B.1. The conclusions presented in the main article could also be reached with a periodic SFH, confirming that the type of SFH does not fundamentally impact the results of the article. To also test if the wavelength range or the type of data (photometric or spectroscopic) could influence the results presented in this paper, we also present the same Mdust vs. Mstar diagram in Fig. C.3: while we do not observe any meaningful differences in the top panel if we do not use the sub-mm data, the two sequences detected in Fig. 16 disappear in the bottom panel, showing that the information from the spectrum is fundamental for identifying the stardust and ISM dust sequences.

thumbnail Fig. C.1.

This figure is the same as Fig. 6. However, for this one we assume a periodic SFH. This alternate version allows to reach the same conclusion than Fig. 6, with a population of GELDAs.

thumbnail Fig. C.2.

This figure is the same as Fig. 16. However, for this one we assume a periodic SFH. This alternate version also allows to reach the same conclusion than Fig. 6, with a population of GELDAs.

thumbnail Fig. C.3.

Test on the stability of the results with the type of data used: Top - This figure is created by fitting the spectrophotometric data, as in Fig. 16 and Fig. C.2, except that we do not use the sub-mm ones. The trend observed in this case is almost identical, which confirms the less important role of the sub-mm data in separating the two parallel sequences. Bottom - This figure is created by only fitting the photometric data, including the sub-mm ones but not the spectroscopic one. The trend observed in this case is different from when we make use of the NIRSpec spectra. We still do see a small decrease at lower stellar mass, even though the less-marked downturn suggests that without spectroscopy, some strong spectral information, and especially the line ratios, is missing. However, the photometric data still bring an information on the dust attenuation because of the UV slope βFUV. The correlation of βFUV with the dust mass is much less significant, and leads to this smaller difference in dust mass, even at low stellar masses which makes the second lower sequence less prominent.

All Tables

Table 1.

Extract from Harikane et al. (2023): possible AGNs in the analysed sample.

In the text
Table 2.

Measured UV slope and fluxes of emission lines of GELDA stacked spectrum are in erg/cm2/s/Å.

In the text
Table B.1.

CIGALE modules and input parameters used for all the fits. BC03 means Bruzual & Charlot (2003), and the Chabrier IMF refers to Chabrier (2003).

In the text

All Figures

thumbnail Fig. 1.

Distribution of spectroscopic redshifts derived by fitting spectrophotometric data with CIGALE. We kept 173 objects with the most robust redshifts. Galaxies are mainly found in two sub-populations at 4 ≲ z ≲ 6 and 5 ≲ z ≲ 6, with a tail extending to z ≲ 12.
In the text
thumbnail Fig. 2.

Comparison of CIGALE fluxes to those measured with LiMe (Fernández et al. 2024). The full results on the CEERS catalogue redshifts and line measurements will be discussed in Arrabal Haro et al. (in preparation). The points corresponding to each emission line are colour-coded to better identify each species (see inside caption). AGN lines are identified by crosses. The most ultraviolet lines (CIV λ1549 Å and HeII λ1641 Å in the lower part of the plot) are not in agreement with those estimated by LiMe as shown by the offset from the solid 1-to-1 line. Instead, we find that other lines are systematically different by about about 20% (dashed line). The observed difference might find an origin in the subtraction of the continuum.
In the text
thumbnail Fig. 3.

Comparison of fluxes computed by CIGALE (Fluxprism) with Gaussian fitting of the lines observed in the grating configuration (Fluxgrating) for objects observed in both configurations. The points corresponding to each emission line are colour-coded to better identify each species (see inside caption). The diagonal line shows the 1-to-1 relation (Fluxprism = Fluxgrating), while the dashed lines present offsets (Fluxprism = 0.5 Fluxgrating) and (fluxprism = 2.0 Fluxgrating). The horizontal and vertical dashed lines show the level of the 3σ background around the [NII] + Hα lines.
In the text
thumbnail Fig. 4.

Distribution of stellar masses derived by fitting spectrophotometric energy distributions of our galaxy sample. In this plot, we only show galaxies with βFUV derived from CIGALE's fits. The range of stellar masses reaches Mstar values as low as a few 107 M⊙. The colour-coding shows that the least massive galaxies clearly have bluer UV slopes (βFUV≤−2.2), while the most massive galaxies have redder slopes (−1.5<βFUV), as expected from the dependence of UV dust attenuation on stellar mass (see e.g. Bogdanoska & Burgarella 2020; Weibel et al. 2024; Bouwens et al. 2016). The vertical numbers show the mean stellar mass for each sample.
In the text
thumbnail Fig. 5.

Star formation rate as function of stellar mass for our galaxy sample. For some galaxies, the observed spectra are not good enough to provide a reliable fit of the spectrum (f λ ∝ λ FUV β $ {}_{\lambda } \propto \lambda ^\beta _{{\mathrm {FUV}}} $) for the UV slope. For those galaxies, we rely on CIGALE's fits. Here, the larger symbols correspond to UV slopes spectroscopically estimated from the observed data, while the smaller symbols are from the CIGALE fits. Blue, green, and red symbols, respectively, show galaxies with blue, intermediate, and red slopes (see inside caption). Whatever the origin of the UV slopes, the lower mass galaxies present a bluer UV slope, while the more massive ones have are redder. In this plot, we superimpose the best-fit MS from Speagle et al. (2014) in the range z = 4–12, colour-coded from magenta to dark red (bottom to top).
In the text
thumbnail Fig. 6.

Far-UV dust attenuation AFUV as a function of stellar mass, log10(Mstar). The symbols are colour-coded and sized with the redshift. The red objects correspond to the highest redshift bin defined in this paper (z > 8.8). The majority of highest-redshift GELDAs (5 out of 6) are found in the low-AFUV part of the plot. This sample of galaxies at z > 8.8 is small, but it might confirm an expected decreasing evolution of the dust attenuation with the redshift. The dashed line is the ‘consensus’ law from Bouwens et al. (2016), while the continuous line is the best linear fit (AFUV vs log10(Mstar)) to the present data. GELDAs are shown with an upside-down Y, and AGNs are shown with crosses.
In the text
thumbnail Fig. 7.

Difference between metallicity estimates (12 + log10(O/H) defined in Eq. (1)): a comparison of the metallicities measured for some of our galaxies in the literature (Nakajima et al. 2022, 2023; Sanders et al. 2024) via auroral lines, with our own estimates. At relatively high metallicities, above 12 + log10(O/H) ∼ 7.4–7.6 (that is Z/Zgas 0.1–11% Z⊙), the differences remain within σ(12 + log10(OH))<0.05. This is about the same dispersion between other metallicity estimates, although on a smaller sample. However, there is a disagreement at lower metallicities, and CIGALE's 12 + log10(O/H) present an offset that might be systematic or increasing with lower metallicities by about σ (log10(O/H)) < 0.10. But the direct method is not well calibrated at log10(O/H) < 7.4–7.6 because there are less than 5 objects with auroral lines at such low metallicities estimated via the direct-method oxygen abundances.
In the text
thumbnail Fig. 8.

Direct comparison of metallicities (12 + log10(O/H) defined in Eq. (1). Left: this panel shows the metallicities derived using various published calibrations as a function of the mean metallicities computed from each measured galaxy. The dashed black line is the 1-to-1 relation. For a given metallicity on an x-axis, we could get a very wide range on the y-axis depending on which line ratio is used. The black dots are the metallicities from CIGALE. The red-orange ones from Nakajima et al. (2023) and the blue-green one from Sanders et al. (2024). The color coding is given in the legend inside the plot. Right: this panel presents the same information with a rolling average. Most of the metallicity values agree within about ±0.5 dex and the metallicities from CIGALE are approximately in the middle range. However, we notice that the ones involving the nitrogen lines lead to much lower metallicities (see text for more details.
In the text
thumbnail Fig. 9.

Evolution of metal-to-stellar mass ratio as function of specific star formation rate in this metal formation rate diagram (MFRD) presented in the same way we defined the dust formation rate diagram (DFRD) showing how the specific dust mass (Mdust/Mstar) changes with the specific star formation rate (SFR/Mstar) in Burgarella et al. (2022). This type of diagram is normalised to Mstar and allows a better comparison between galaxies. The symbols are colour-coded with the redshift. The pink shaded area presents the 2-σ confidence interval for the distribution in MZ/Mstar, and the solid black line shows the rolling average. We observe a decreasing trend with decreasing sSFR. The highest redshift galaxies are located in the lowest part of the plot.
In the text
thumbnail Fig. 10.

Correlation of estimated dust mass, Mdust, with dust attenuation, A(Hα). The most relevant parameter to predict the dust mass is the dust attenuation A(Hα) derived from the Hα/Hβ Balmer decrement. The correlation with A(Hα) alone accounts for 76% of the variation in Mdust. The spectral information brought by NIRSpec is thus fundamental to estimating the dust masses. Blue, green, and red symbols, respectively, mean Mdust≤105 M⊙, 105 < Mdust≤106 M⊙, and Mdust > 106, while magenta circles show GELDAs.
In the text
thumbnail Fig. 11.

Same colours as in Fig. 10. Correlation of the estimated dust mass, Mdust with the UV slope, βFUV. The correlation with βFUV alone accounts for 44% of this variation, which confirms that the spectral information on lines is the most important one. The other tested parameters – metallicity (about 2%) and redshift (<1%) – as well as the level of the sub-millimetre upper limits, are not significantly correlated with Mdust in this analysis.
In the text
thumbnail Fig. 12.

Use of mock analysis to estimate minimum dust mass that we can actually estimate using our approach. Top: x-axis shows modelled Mdust computed from the best-fit models and for each of our galaxies. CIGALE is able to recover these input dust masses (y-axis) by fitting the mock data down to about log10(Mdust) = 5.0. Parts of the objects in green and all the objects in blue should thus be considered as upper limits. These objects that are the ones identified as transitioners between the upper sequence and a possible lower sequence are well below the prime sequence (red dots). Bottom: log10(Mdust) versus log10(Mstar) diagram showing that the lowest dust masses correspond to the lowest stellar masses. However, for the same stellar-mass range, we do observe a wide range of dust masses. Normal, mid, and low Mdust mean Mdust≤105 M⊙, 105 < Mdust≤106 M⊙, and Mdust > 106 M⊙, respectively.
In the text
thumbnail Fig. 13.

Evolution of metal and dust masses. MZ/Mstar (blue dots) and Mdust/Mstar (red dots) both regularly increase with the age of the Universe (top); more metals and dust grains are formed as the Universe ages. We note that even for galaxies (blue symbols) in the early Universe (ageUniverse≲600 Myr or z ≳ 9), MZ/Mstar never goes below a few 10−3, possibly suggesting a fast rise of metals that produces the observed threshold. MZ/Mstar increases faster than Mdust/Mstar, with a much larger dispersion for Mdust/Mstar. The light blue and red areas show the mean and 2σ confidence interval of the distribution within several bins. At the bottom, same as aboce a function of sSFR. MZ/Mstar (in blue) and Mdust/Mstar (in red) decrease from high to low sSFR (bottom), as also shown in, e.g. Palla et al. (2024) and Shivaei et al. (2022). We also note a larger dispersion at lower sSFR for Mdust/Mstar, but only for Mdust/Mstar.
In the text
thumbnail Fig. 14.

Stacked spectrum of objects selected in lower sequence. Vertical lines show the location of a few usual emission lines (see inside caption). We show the position of the lines, including where the HeI lines would be expected. None of them are detected, confirming these objects do not contain a dominant AGN. This stacked spectrum is similar to that of a young starburst, with a blue UV slope and prominent hydrogen and [OIII) emission lines.
In the text
thumbnail Fig. 15.

SFR surface density as function of gas-mass surface density for a sample of local and high-redshift galaxies, including HZ10, LBG-1, AzTEC-3, and CRLE, adapted from Pavesi et al. (2019) and Shi et al. (2011). Our objects, GELDAs, and non-GELDAs are very dense, and all of them are found slightly below other objects, which might be related to the very high gas density in these objects where the turbulence can produce a negative effect on the SFE, which is thus lower than in other high-redshift galaxies at the same gas surface density. From bottom to top, the shaded areas correspond to low-surface-brightess (magenta), late-type (blue), early type (green), hi-z star-forming (yellow), and hi-z sub-millimetre (red) galaxies. They are also given inside the figure.
In the text
thumbnail Fig. 16.

Mdust as function of Mstar. Top: Dots are colour-coded according to redshift. The size of the symbols provides us with information on 12 + log10(O/H), with the largest sizes corresponding to the largest metallicities. We show density contours as heavy black lines. At log10 (Mstar) ∼ 108−9 M⊙, we observe a transition from an apparent high sequence to a lower one. The upper sequence is similar to that observed at low redshift (e.g. Beeston et al. 2018). The lower objects have not been identified before in this (Mdust vs. Mstar) diagram, except for a few objects (green crosses, De Vis et al. 2017). They share the same location in the plot and are extracted from a sub-sample of galaxies in an HI-selected local-galaxy sample from GAMA and H-ATLAS. Bottom: Models plotted on density contours. Hydrodynamical code dustygadget (Graziani et al. 2020) is in yellow, a cosmological hydrodynamical simulation coupled with a chemical evolution model (Mancini et al. 2015) is in green, and a suite of cosmological, fluid-dynamical simulations of galaxies (Esmerian & Gnedin 2024) are in pink. The upper brown line (Witstok et al. 2023) shows where galaxies with ISM-grown grains should be. The bottom orange line corresponds to galaxies with only stardust (Witstok et al. 2023) that underwent a 95% destruction of grains by reverse SNae shock (Witstok et al. 2023). The objects observed at the bottom of the top panel correspond to the transition between galaxies only containing stardust to galaxies dominated by ISM dust.
In the text
thumbnail Fig. 17.

Objects colour-coded by redshift. AGNs are shown as crosses, and GELDAs are upside-down Ys. This figure shows that GELDAs (both z ⩾ 8.8 in red and z < 8.8 in blue) have lower metal masses than the rest of the sample. This would support the hypothesis that the origin of dust grains in these objects might not be due to ISM growth.
In the text
thumbnail Fig. 18.

Distribution of metallicity for GELDAs and non-GELDAs. The left panel presents the kernel density estimation (KDE) with an Epanechnikov kernel and a bandwidth = 0.5 both for GELDAs (solid blue line) and non-GELDAS (solid red line). The dashed black lines show the differences. The blue and red shaded areas show where GELDAs and non-GELDAs are the dominant population, respectively. We clearly see that GELDAs preferentially cluster at low metallicity, with a threshold estimated at 12 + log10(O/H) = 7.60. The Kolmogorov-Smirnov test confirms that the difference is highly significant. We show predictions for the transition from stardust to ISM from some of the models (see main text for more works): Asano et al. (2013) (orange shaded area), Zhukovska (2014) (green shaded area), and Feldmann (2015) (dashed orange line). Rémy-Ruyer et al. (2014) found a larger scatter in their dust-to-gas ratio covers almost the same metallicity range: 7.2 ≲ 12 + log10(O/H) ≲ 8.5. All of these metallicities are in good agreement with our estimated threshold. The middle and right panels show the rolling averages (window size = 15) of the metallicity as a function of AFUV and log10 (Mdust). Both panels confirm the difference between GELDAs (blue) and non-GELDAS (red).
In the text
thumbnail Fig. A.1.

This flag must be set to ”True” in the pcigale.ini file to fit spectroscopic data.
In the text
thumbnail Fig. B.1.

a) Objects in the upper sequence (larger Mdust) in Mdust vs. Mstar plot. We present a sample of spectral fits over the whole spectral range, that is including NIRSpec spectroscopy and the sub-millimeter data (mostly upper limits).
In the text
thumbnail Fig. B.2.

b) Same as Fig. B.1 for the same objects but only the fits of the NIRSpec spectrum is shown, which is a zoom in the previous plots.
In the text
thumbnail Fig. B.3.

Same as previous Fig. B.1.
In the text
thumbnail Fig. B.4.

Same as Fig. B.2.
In the text
thumbnail Fig. B.5.

Objects in the lower sequence (lower Mdust) in Mdust vs. Mstar plot. Same as Fig. B.1 but for objects in the lower sequence in the Mdust vs. Mstar plot.
In the text
thumbnail Fig. B.6.

Same as Fig. B.5 for the same objects but only the fits of the NIRSpec spectrum is shown, which is a zoom in the previous plots.
In the text
thumbnail Fig. C.1.

This figure is the same as Fig. 6. However, for this one we assume a periodic SFH. This alternate version allows to reach the same conclusion than Fig. 6, with a population of GELDAs.
In the text
thumbnail Fig. C.2.

This figure is the same as Fig. 16. However, for this one we assume a periodic SFH. This alternate version also allows to reach the same conclusion than Fig. 6, with a population of GELDAs.
In the text
thumbnail Fig. C.3.

Test on the stability of the results with the type of data used: Top - This figure is created by fitting the spectrophotometric data, as in Fig. 16 and Fig. C.2, except that we do not use the sub-mm ones. The trend observed in this case is almost identical, which confirms the less important role of the sub-mm data in separating the two parallel sequences. Bottom - This figure is created by only fitting the photometric data, including the sub-mm ones but not the spectroscopic one. The trend observed in this case is different from when we make use of the NIRSpec spectra. We still do see a small decrease at lower stellar mass, even though the less-marked downturn suggests that without spectroscopy, some strong spectral information, and especially the line ratios, is missing. However, the photometric data still bring an information on the dust attenuation because of the UV slope βFUV. The correlation of βFUV with the dust mass is much less significant, and leads to this smaller difference in dust mass, even at low stellar masses which makes the second lower sequence less prominent.
In the text

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.