Finding dusty AGNs from the JWST CEERS survey with mid-infrared photometry

JWST Cosmic Evolution Early Release Science (CEERS, Finkelstein et al., 2017; Yang et al., 2023a) survey provides a continuous observation from near-infrared (NIR) to mid-infrared (MIR) wavelength by JWST Near Infrared Camera (NIRcam), Mid-Infrared Instrument (MIRI, Rieke et al., 2015; Rigby et al., 2023), and Near-Infrared Spectrograph (NIRSpec).

In this work, we focus on 4 MIRI pointings in the Extended Groth Strip (EGS) legacy field with 6 continuous broad-band filters (F770W, F1000W, F1280W, F1500W, F1800W, and F2100W). Two pointings have all 6 bands (observation ID: o001_t021 and o002_t022), while the other two pointings have only 4 bands, missing F770W and F2100W (observation ID: o012_t026 and o015_t028). We use photometry reported in Wu et al. (2023) which conducts a series of processes including 80% completeness measurements and comparisons with previous model predictions (e.g., Gruppioni et al., 2011; Cowley et al., 2018). Besides, given that STScI has released the new MIRI flux calibration (see https://www.stsci.edu/contents/news/jwst/2023/updates-to-the-miri-imager-flux-calibration-reference-files), we have followed the manuscripts to update the photometry as well. The related updates are 15.0%, 4.1%, 0.8%, 1.2%, 3.3%, and 3.3% higher than previous for F770W, F1000W, F1280W, F1500W, F1800W, and F2100W, respectively. Wu et al. (2023)’s 80% completeness flux in CEERS survey are also scaled for F770W ($0.28\mu Jy$), F1000W ($0.65\mu Jy$), F1280W ($1.26\mu Jy$), F1500W ($2.0\mu Jy$), F1800W ($5.0\mu Jy$), and F2100W ($13.4\mu Jy$), respectively. More technical details are described in Wu et al. (2023).

To study the physical properties of AGNs, it is insufficient to utilise only MIR photometry. Therefore, a photometric catalogue covering multiple wavelengths is necessary. CANDELS-EGS Multi-wavelength catalogue (Stefanon et al. 2017) constructed from the Cosmic Assembly Near-IR Deep Extra-galactic Legacy Survey (CANDELS: Grogin et al., 2011; Koekemoer et al., 2011) offers a broad observation range from near-UV to MIR in the EGS field. By matching our sources with CANDELS-EGS, we obtain a 21 multi-band catalogue with 15 bands from CANDELS-EGS and 6 bands from CEERS MIRI pointings (Ling et al. 2024, see Table 1 for details). Our catalogue includes 573 matched sources, each with at least one MIRI detection.

Code Investigating GALaxy Emission v2022.1 (CIGALE; Boquien et al., 2019) is a Python-based code for fitting spectral energy distributions (SEDs) and can estimate the physical properties of galaxies with observations from far-UV to radio emission. It provides various modules such as star-forming history, dust attenuation, dust model, and AGN model to fit observations. In this work, we mainly follow the configuration from Yang et al. (2023b). We slightly change the number of AGN fractions, the dust attenuation model, and the given redshift. In Table 3, we list the details for the full parameters we adopt.

Note that in CIGALE, AGN contribution is called AGN fraction. Other literature might treat AGN fraction as AGN number fraction (e.g., Gu et al., 2018), which is another main discussion in this work. To avoid confusion, here, we use the term AGN contribution to replace the AGN fraction.

AGN contribution ($f_{\rm AGN,IR}$) is the ratio between AGN luminosity ($L_{AGN}$) and total infrared luminosity, describing how much the AGN emission contributes to the total emission. The definition of it is described in equation 1:

where IR is defined in $8-1000\mu m$ by Kennicutt 1998. The default wavelength range to define the AGN contribution in CIGALE is $8\,\mu m-1000\,\mu m$. Note that the total IR luminosity estimated by CIGALE might have uncertainty due to a lack of far-infrared (FIR) observations in this work. The relative problems will be discussed in §3.1. Additionally, since our sample selection criterion of the AGN contribution is $f_{\rm AGN,IR}\geq 0.2$ (described in §2.3), we add extra choices of AGN contributions of 0.13, 0.15, and 0.18 during SED fitting to improve the accuracy of SED fittings near our threshold. The AGN contributions can take any of the following values during the SED fitting, from 0.0 to 0.99 as shown in Table 3.

For the dust attenuation, we follow the same parameter set in Wang et al. (2020) instead of the one in Yang et al. (2023b). Yang et al. (2023b) reports that CIGALE may underestimate the redshift compared with spec-$z$ from CANDELS-EGS catalogue. However, to pursue an accurate estimation, we directly use spec-$z$ from the CANDELS-EGS catalogue if available ($\sim$23% of our sample). For the remaining no-spec-$z$ sources, we give a range of redshift to allow CIGALE to estimate photo-$z$.

CIGALE provides a Bayesian estimation for the output of physical properties. This estimation considers all possible models to weigh a statistically-based physical property, while the best-fit SED output is based on the minimal $\chi^{2}$ of the model. Statistically speaking, the Bayesian output is more reliable than the best fit. Yang et al. (2023b) also highlights the robustness of the Bayesian output. Therefore, in this paper, we choose the Bayesian output for the analysis rather than the best fit.

According to the Bayesian $f_{\rm AGN,IR}$, we check the estimation quality. Fig. 2 displays the results. The median $f_{\rm AGN,IR}$ and error for all sources are $\sim$0.068 and $\sim$0.075, respectively. The error is not too large to affect our following analysis significantly as our selection criterion is $f_{\rm AGN,IR}\geq 0.2$, making our selection less likely to be severely contaminated by SFG.

Fig. 1 displays the performance of SED fittings with their MIRI cutouts.

After our initial run of CIGALE on 573 sources we select our initial sample based on the SED fitting results.

First, our work mainly focuses on using mid-IR photometry to select AGNs, Yang et al. (2023b) provides a criterion that each source should be detected by at least 2 bands in JWST MIRI. However, to obtain a comprehensive AGN census and ensure that the MIR colour-colour diagram (e.g., Kirkpatrick et al., 2017) is available, we impose a more stringent constraint, requiring at least 3-band detection.

Additionally, since two of the pointings (observation ID: o012_t026 and o015_t028) miss 7 $\mu m$ and 21 $\mu m$ detection, we are unable to apply the 7 $\mu m$ 80% completeness flux limit, which has the faintest flux limit (Ling et al., 2022; Wu et al., 2023). Therefore, we adopt the second faintest 10 $\mu m$ 80% completeness flux limit (Ling et al., 2022; Wu et al., 2023) to exclude sources that fall below the flux limit. Some sources might have no 10 $\mu m$ detection by MIRI, so we substitute the estimated 10 $\mu m$ flux from CIGALE for undetected flux to make the completeness selection feasible and doable.

Finally, we check the performance of SED fittings and find out that there are 2 stars in the sample. In addition to stars, we find a source whose $\chi^{2}$ is 9.6, which is larger than the median $\chi^{2}$ of the rest of the sample ($\sim$1.14). After checking the image from the Hubble Space Telescope (HST), Spitzer, and JWST, this source shows no abnormality and is included in our final sample. In summary, we exclude 2 stars. 253 sources remain in the final sample. Fig. 3 shows the flowchart of our sample selections.

To define our sources as the candidates, we follow the same criterion as Wang et al. (2020), requiring $f_{\rm AGN,IR}\geq 0.2$ for the candidates. We choose this criterion to enable a proper comparison with Wang et al. (2020) and to demonstrate the unprecedented sensitivity of JWST (e.g., Kim et al., 2021; Ho et al., 2021). Additionally, We divide our candidates into two based on the following criteria: composite ($0.2\leq f_{\rm AGN,IR}<0.5$) and AGN ($f_{\rm AGN,IR}\geq 0.5$). Among 42 candidates, 30 of them are composites and 12 of them are AGNs. The remaining 210 sources are classified as SFGs because of low AGN contributions ($f_{\rm AGN,IR}<0.2$).

For the discussion on $f_{\rm AGN,IR}$ in this work, the absence of FIR observations is a problem for SED fittings. The uncertainties may occur when the scientific results are analysed by physical properties estimated from SED fittings. To explore this influence, it is necessary to apply mock observations. In CIGALE, a mode called "savefluxes" can introduce a simulation to generate mock observations. We utilise the savefluxes model to generate 100 observations to analyse how significant the uncertainty is when the absence of FIR observations occurs. Table 4 lists the parameters applied to generate the mock observations. We set the majority of parameters to be one choice and merely change the AGN contribution and redshift since we focus on two crucial physical quantities in this work: AGN contribution and total IR luminosity to constrain the mock performance. Both quantities are strongly affected by the choice of AGN contribution and redshift parameters in CIGALE. Note that we set the redshift range from 0.5 to 1.0 rather than starting from 0.01 due to the avoidance of overflow in CIGALE if the redshift is smaller than 0.5.

We separate the mock observations into two groups. One group (hereafter, the group with-Herschel) contains the Spectral and Photometric Imaging Receiver (SPIRE, Griffin et al., 2010) PSW (250$\mu$m), PMW (350$\mu$m), and PLW (500$\mu$m) in the Herschel Space Observatory (HSO). The other group (hereafter, the group without-Herschel) does not. For the flux errors of the SPIRE mock observations, we adopt the $\rm SNR=40$ flux-error relation from Pokhrel et al. (2016) to calculate the errors. For the other bands, we calculate the error-flux ratio for each band in each source. Fig. 4 shows the distribution of the error-flux ratio for JWST as an example. We take the median value to generate the errors for each band.

Next, we utilise the parameter list in Table 3 to re-fit the SEDs for both groups and analyse their performance respectively. Fig. 5 shows their performance. Note that there are some small point displacements in both figures, this phenomenon is due to the choice of using Bayesian estimations for AGN contribution and total IR luminosity. Bayesian estimations can consider more situations than using input numbers. The estimated values for both groups are consistent according to the estimated values (for without-Herschel, $\sim$2.7% smaller for AGN contribution and $\sim$7.5% smaller for total IR luminosity) and the fitting slopes (1.031 for AGN contribution and 0.96 for total IR luminosity). However, for the error estimations, the median error of AGN contribution for the group with-Herschel (without-Herschel) is $\sim$0.065 ($\sim$0.104); The median error of total IR luminosity is $\sim 9.7\times 10^{9}$ ($\sim 1.47\times 10^{10}$) in solar luminosity $\rm L_{\odot}$. Fig. 6 shows detailed results, proving that the without-Herschel group estimates larger errors than the with-Herschel group ($\sim$65.2% larger for AGN contribution and $\sim$52.1% larger for total IR luminosity, both take the median value of the fraction in Fig. 6). It indicates that the catalogue in which FIR observations are absent surely enlarges the uncertainty of estimated physical properties. In conclusion, although the uncertainty of errors is enlarged, we claim that lacking FIR observations in this work does not largely and significantly change our conclusions. Additionally, compared with the with-Herschel group, the estimated total IR luminosity in the without-Herschel group seems to be slightly underestimated ($\sim$7.5%). The majority of points are below the ratio = 1.0. This underestimation might refer to the importance of FIR observations as well.

We start to discuss our fitting obtained through CIGALE. In the following subsections, we concentrate on two essential properties: AGN contribution and AGN number fraction. To establish the relations of AGN contribution and AGN number fraction with other physical properties, we have depicted a preliminary picture of our sources distribution in Fig. 7. From Fig. 7, our AGN candidates are mostly located in higher redshift with higher AGN contribution rather than the composite candidates, which is consistent with the results in Toba et al. (2020). However, it is not sufficient to claim this relation. An advanced analysis is necessarily introduced. To determine whether there are relations between AGN contribution and total IR luminosity more specifically, we investigate both AGN contribution and AGN number fraction in §3.3 and §3.4, respectively, to obtain a clearer understanding of their relations.

In addition, we compare our candidates with those of Yang et al. (2023b), who, before our work, also utilised the same field and sample to select AGN candidates. However, their criteria were $0.1\leq f_{\rm AGN,IR}<0.5$ for composite sources and $f_{\rm AGN,IR}\geq 0.5$ for AGN. Applying these criteria to our candidates, the number of composites increases to 83, while the number of AGN remains at 12, yielding number ratios of $83/253=0.33\pm 0.04$ and $12/253=0.047\pm 0.014$, respectively. Yang et al. (2023b) identified 102 composite sources and 25 AGN from 560 sources, resulting in $0.18\pm 0.02$ and $0.045\pm 0.009$ ratios for composites and AGN, respectively. While the AGN ratios are consistent, our composite ratio is higher.

The exact reason for this difference is unclear, but possible sources include the fact that Yang et al. (2023b) used sources detected in at least two MIRI filters, whereas we required detections in three filters. This results in a smaller sample size for our analysis (253 vs. 560), which might lead to the selection of intrinsically brighter sources and a slightly larger fraction of composites. Furthermore, as shown in Fig. 2, the number of objects changes drastically with small variations in $f_{\rm AGN,IR}$, which can occur due to slight differences in the fitting parameters between our analyses. This could also contribute to the observed difference in the fraction of composite sources.

We calculate the average AGN contribution in each bin and check whether the relation between AGN contribution and redshift or total IR luminosity is clear or not. Note that we ignore the bins which have only one source for consistency and proper statistical meaning.

Fig. 8 illustrates AGN contribution as a function of redshift. The results in Wang et al. (2020) are depicted as well to show the comparison. It seems that the green and the orange group are increasing. However, if the p-value is considered, both exceed the p-value threshold ($p\leq 0.05$), which means that we cannot conclude a tight relation between AGN contribution and redshift in this work statistically. This discrepancy with Wang et al. (2020) might be due to the lack of sources. Therefore, a larger sample size is required, enabling us to study the relationship more exhaustively.

Fig. 9 is AGN contribution as a function of total IR luminosity. It seems that AGN contribution is decreasing with increasing luminosity, especially for the purple and blue groups. This result is inconsistent with Chiang et al. (2019). Chiang et al. (2019) reports that AGN contribution is increasing as the luminosity goes higher. Such difference might be due to the source distribution of the sample in this work. The majority of sources ($\sim$69.8%) are below $10^{11}L_{\odot}$, which are not the typical luminous infrared galaxies (LIRGs, Sanders & Mirabel, 1996). We claim that this reason might be the key to link to the decreasing result. Besides, all groups extend their tails to their faint ends compared with Wang et al. (2020). This extension might be attributable to the level of sensitivity between AKARI and JWST. JWST has a high sensitivity that can capture many fainter galaxies than before, inferring the power of finding fainter AGNs in the Universe. This achievement can allow us to investigate the evolution of AGNs and even the reionization of the Universe (Giallongo et al., 2015; Grazian et al., 2018) further. On the other hand, Fig. 9 might indicate that JWST has successfully accomplished our expectations.

AGN number fraction ($f_{\rm num}$) can demonstrate how AGN populates in each redshift bin. The definition of AGN number fraction is presented in eq. 2:

where $N_{\rm Candidate}$ is the number of our composite candidates plus AGN candidates.

Fig. 10 shows our result. We compute the AGN number fraction in each bin and utilise the 1$\sigma$ errors for small statistics presented in Gehrels (1986) to define the error bar for each bin. Wang et al. (2020) reports that AGN number fraction might have an increasing relation as a function of redshift. According to our results, only the orange group displays a low-p-value ($p\leq 0.05$) increasing relation. Nevertheless, the small sample size and small statistics make the uncertainty large enough to be unable to identify the truth of the trend. Kirkpatrick et al. (2023) reports that the AGN population is also increasing. However, their sample is also affected by the small statistics, making the uncertainty hard to constrain. In contrast, Hashiguchi et al. (2023) utilises much larger ($\sim 27000$) sources to investigate the AGN number fraction in galaxy clusters. Their sample size is large enough to conclude a tightly increasing AGN number fraction against the redshift. To overcome the lack of sources, a larger mid-IR survey is necessary to increase more sources, lowering the difficulty of identifying scientific relations. However, the good news is that thanks to the high sensitivity of JWST, we can unearth more high-$z$ galaxies that are difficult to observe by previous telescopes. Focusing on the blue, green, and orange groups, we conclude that we successfully extend and reveal the results in Wang et al. (2020) to the higher redshift. Besides, the result shows the difference between two different fields as well. 4 MIRI pointings ($\sim$ 8 arcmin²) are much smaller than the NEP-wide field ($\sim$ 5.4 deg²) in Wang et al. (2020), the field of view limits the size of our sample in 4 MIRI pointings. This might be a key reason for our result.

In addition to redshift, we check AGN number fraction as a function of total IR luminosity shown in Fig. 11. Our result seems to have no obvious trends in different groups, which are consistent with Wang et al. (2020). Furthermore, none of the p-values for each group are smaller than 0.05, leading us to difficulty identifying the existence of trends between AGN number fraction and total IR luminosity. However, if we focus on the groups except for the orange group, these groups show higher AGN number fractions at each fainter end. Nevertheless, Kartaltepe et al. (2010) utilises Spitzer to investigate AGNs, reporting that the AGN number fraction is increasing at the ULIRGs and Hyper-luminous infrared galaxies (HyLIRGs ($L_{\rm TIR}\geq 10^{13}L_{\odot}$), related work can be found in Gao et al., 2021). Chiang et al. (2019) concludes the same trend by AKARI as well, reporting the luminous sources are dominated by the AGN population. These discrepancies might be due to the sensitivity difference between the previous telescope and JWST, causing the previous works to be limited to low sensitivity. In this work, with the higher sensitivity of JWST, we can extend our understanding down to fainter galaxies ($L_{\rm TIR}<10^{11}L_{\odot}$), not just LIRGs or ULIRGs. Moreover, as we mentioned above, the majority of the luminosity of our sources ($\sim$69.8%) is not LIRGs. Consequently, combining the results in Fig. 9 and Fig. 11, they both suggest that we might successfully capture less luminous AGNs, which is a key goal (capturing fainter AGNs) in this work. This key result can lead us to believe that JWST can unearth more faint AGNs in the distant Universe (e.g., Harikane et al., 2023). The majority of sources in Wang et al. (2020) are LIRG and ULIRG due to the limitation of sensitivity for AKARI, and Kirkpatrick et al. (2023) reports the discovery of the population of mid-IR weak galaxies (where the mid-IR luminosity is mostly $\leq 10^{10}L_{\odot}$) in the high-z Universe. Combining with the results in this work, we can conclude that JWST truly reveals a new era of faint AGN research.

In this section, we briefly discuss the X-ray-selected AGNs in our sample. An X-ray survey called AEGIS-X Deep (AEGIS-XD) survey by Chandra X-ray observatory (Nandra et al., 2015) provides a wild observation to investigate the properties of X-ray sources in the Universe, the CEERS field is within the survey area as well.

Among our sample, AEGIS-XD has 13 matched X-ray sources. Adopting the X-ray AGN criterion by X-ray luminosity $L_{\rm X}\geq 10^{43}$ erg/s (Auge et al., 2023), we have 6 X-ray-selected AGN candidates and 4 of them are inside the candidates in this work ($\sim$66.6%). The remaining 2 sources are classified as SFGs by our selection criteria, moreover, their AGN contributions are minimal ($f_{\rm AGN,IR}\leq 0.01$), which is quite lower than our criterion. Apart from the X-ray-selected AGNs, we still have 38 candidates not detected by the AEGIS-XD survey. However, with the MIRI on JWST, we complement the gap of using X-ray to select AGNs.

Besides, we investigate the relation between X-ray luminosity and total IR luminosity. Fig. 12 illustrates the relation, a slightly increasing trend is clear. To confirm whether the relation is statistically reasonable, we apply the p-value test for X-ray and total IR luminosity. The result shows $\rm p=0.928$, which is larger than 0.05, inferring that the relation in this work is statistically insignificant.

Here we adopt the JWST MIRI colour-colour diagram for $z\sim 1$ ($0.75\leq z<1.25$), $z\sim 1.5$ ($1.25\leq z<1.75$), and $z\sim 2$ ($1.75\leq z<2.25$) from Kirkpatrick et al. (2017) and compare to its performance.

For the flux density ratio, we utilise the estimated flux value from CIGALE to replace the undetected flux for the feasible use of the colour-colour diagram. Kirkpatrick et al. (2017) provides different criteria to select these three objects: $f_{\rm AGN,MIR}<0.3$ as SFGs, $0.3\leq f_{\rm AGN,MIR}<0.7$ as composites, and $f_{\rm AGN,MIR}\geq 0.7$ as AGNs. Note that the aforementioned AGN contribution in §2.2 is defined in $8-1000\mu m$ for total IR luminosity. Kirkpatrick et al. (2017) uses AGN mid-IR contribution ($f_{\rm AGN,MIR}$) which is defined in $5-15\mu m$ for mid IR luminosity. In Table 3, we do not particularly set the wavelength range of AGN contribution to be mid-IR range. Therefore, we integrate the mid-IR luminosity for dust emission and AGN emission from the best SED fitting result of our sample respectively, and calculate AGN mid-IR contribution using the same equation form as eq.1.

Fig. 13, Fig. 14, and Fig. 15 illustrate the results of our sample for both AGN mid-IR contribution and AGN contribution. The majority of our candidates are successfully captured by the colour-colour diagram. However, Kirkpatrick et al. (2023) recently reported that this selection can be contaminated by SFGs and mid-IR weak galaxies which can pretend to be the AGN candidates or have unknown redshift. Therefore, Kirkpatrick et al. (2023) provides two new colour-colour selections (hereafter, MIR-only and NIR+MIR, respectively) to avoid these influences. We apply both methods to select the AGN candidates (hereafter, the cc-AGNs) and compare them with our candidates (hereafter, the SED-AGNs). Fig. 16 shows both results, having 32 (32) cc-AGNs for MIR-only (NIR+MIR) selection. The number of overlapped candidates with the SED-AGNs is 10 (15), showing some missed candidates in both the cc-AGNs and SED-AGNs. We check the missed candidates the SED-AGNs lose and the cc-AGNs lose. The SED-AGNs miss 22 (17) candidates and the cc-AGNs miss 32 (27) candidates.

Given the missed candidates, we check whether they are SFGs for each set of candidates. We focus on possibly the most problematic cases, six SED-AGNs that both cc-selections classified as SFGs. We crosscheck the SED results and the AGN contributions for the 6 candidates. These 6 SED-AGN candidates missed by the cc-selections exhibit small AGN contributions close to our selection criterion ($f_{\rm AGN,IR}\sim 0.2$). Table 2 displays the detailed values and the relative errors. The errors are also large, while the best-fit AGN contributions for each source are also small at almost 0.0. As mentioned in §2.2, Bayesian $f_{\rm AGN,IR}$ is considered more robust than the best-fit $f_{\rm AGN,IR}$ statistically. With this consideration, the cc- and SED-AGNs are not overly inconsistent.

According to the above results, the cc-AGNs lose more candidates than the SED-AGNs, especially in the MIR selection. Kirkpatrick et al. (2023) mentions that the reliability of the NIR+MIR selection is higher than the MIR-only selection, which is consistent with our results. Additionally, Kirkpatrick et al. (2023) points out that applying the colour-colour selection in the small survey is not an efficient way to find AGNs, which agrees with our result of using MIR colour and NIR+MIR colour selection. Therefore, a larger survey (e.g., COSMOS-Web, Casey et al., 2023) is more suitable and might be a potentially useful data to apply colour-colour selections.