Beyond Traditional Diagnostics: Identifying Active Galactic Nuclei with Spectral Energy Distribution Fitting in DESI Data

Introduction

Supermassive black holes with masses of more than a million times that of the Sun (e.g. [81]), are believed to inhabit the centers of galaxies, including the Milky Way (e.g., [54]; [53]). Despite the key role supermassive black holes are thought to play in galaxy formation and evolution, their origin and impact on host galaxies are still a major challenge of modern astrophysics. When supermassive black holes are actively accreting as active galactic nuclei (AGN), they are brighter and easier to detect and thus serve as a proxy to study supermassive black hole properties. Thus, identifying AGN is crucial for understanding galaxy evolution, black hole growth, and cosmic feedback processes (see Harrison & Ramos Almeida [64] for a review).

Despite their important role, a uniform method to identify AGN is challenging as different selection techniques lead to different samples that may not even overlap with each other. AGN are usually identified based on: i) their optical emission lines (e.g., [5]; [79]; [77]), and variability (e.g., [111]) ii) X-ray emission (e.g., [13]), iii) mid-infrared (MIR) emission (e.g., [83]; [133]; [73]; [132]; [4]; [82]; [70]), and variability (e.g., [7]), and iv) radio emission (e.g., [66]; [110]; [134]). Optical spectroscopic selection identifies AGN by detecting broad (BL-AGN, i.e. type 1 AGN) or narrow (NL-AGN, i.e., type 2 AGN) emission lines, indicative of high-energy processes near the black hole (e.g., [5]). However, star formation elevated by a diffuse ionized gas, photoionization by hot stars, metallicity (as indicated by strong forbidden lines such as [N II]$\lambda$6583, [S II]$\lambda$6716, 6731, and [O III]$\lambda$5007), or shocks can be misclassified as AGN (e.g., Milewski et al. 2018). X-ray selection is highly effective because AGN typically exhibit strong X-ray emission that is less likely to be contaminated by star formation processes. However, X-ray surveys are often limited by sensitivity and coverage and may miss obscured AGN (e.g., [55]; [18]; [93]). MIR selection uses color criteria to distinguish AGN from star-forming galaxies due to the characteristic dust emission heated by the AGN ([133]; [4]). The high-energy radiation emitted from the accretion disk heats up the surrounding dust, causing it to emit infrared radiation, particularly in the MIR regime (5 – 30μm). The MIR emission has already proven its usefulness in AGN selection, though it is biased towards the selection of AGN that dominate over the emission from the host galaxies. Radio selection relies on the strong synchrotron emission from relativistic jets, but this method is biased towards radio-loud AGN, which are a minority of the AGN population. Moreover, radio selection can pick up star-forming regions with strong radio emission [110]. Summarizing, accurate AGN identification is often challenging, as each of these common methods captures different subsets of the AGN population and is subject to its own limitations and biases.

The aforementioned limitations of traditional selection methods highlight the need for a multi-wavelength approach to reliably and comprehensively identify AGN. The advent of large multi-wavelength surveys triggered using SED fitting to constrain AGN and their host galaxy properties for statistical samples. Recently, the SED fitting approach revealed the potential not only to derive reliable properties of AGN and host properties (e.g., [92, 104, 16, 9]) but also to identify AGN based on their multi-wavelength information (e.g., [135, 9, 143, 114]). AGN SED modeling techniques are also used as the base for the target selection of forthcoming wide-field spectroscopic surveys such as 4MOST ([96]) and VLT-MOONS ([89]).

In this paper, we identify AGN in the DESI survey using the CIGALE SED fitting code, modeling optical to MIR photometry for millions of sources across diverse types and redshifts. In particular, DESI has already revealed unprecedented samples of dust-reddened QSO ([51]) or changing-look AGN ([59, 58]). DESI value-added catalog (VAC) of AGN selected based on their multi-wavelength information (Juneau et al. in prep.) provides so far the largest AGN spectroscopic sample². This will allow for a more detailed study of the role of AGN in galaxy evolution.

Our study explores the potential of SED-based AGN identification and assesses the necessity of MIR data in minimizing the misclassification of star-forming galaxies as AGN. By integrating multi-wavelength data, we aim to contribute to a more accurate and complete catalog of AGN in the DESI survey. The structure of the paper is as follows. In Sect. 2.1, we provide an overview of the DESI data. In Sect. 3, we describe the AGN selection using different standard techniques. Results with discussion are presented in Sect. 4. Finally, Sect. 5 summarizes the efficiency of SED-based AGN identification.

Throughout this paper, we assume WMAP7 cosmology ([80]), with $\Omega_{\mathrm{m}} = 0.272$ and $H_0 = 70.4$. We also consider the photometry in AB magnitudes ([108]).

DESI data

Overview of DESI

DESI is a robotic, 5000-fiber multiobject spectroscopic surveyor operating on the Mayall 4-meter telescope at Kitt Peak National Observatory (DESI Collaboration et al. 2022). DESI is capable of obtaining simultaneous spectra covering a range of 3600–9800 Å with a resolution of $R = 2000$–5500 of almost 5000 objects over a $3\,\mathrm{deg}^2$ field (DESI Collaboration et al. 2016a,b; Miller et al. 2023; Poppett et al. 2024). DESI is currently conducting a five-year survey to observe approximately 36 million galaxies ([62, 116, 145]) and 3 million quasars (Chauhan et al. 2023) over 14000 $\mathrm{deg}^2$ ([36] (DESI Collaboration et al. 2025)). This campaign will provide ten times more spectra than the SDSS ([144, 2]) sample of extragalactic targets and substantially deeper than prior large-area surveys ([36]). [36] will provide the most precise measurement of the expansion history of the universe ever obtained ([87]).

DESI started to operate in December 2020, with a 5-month Survey Validation (SV; [37]) before the start of the Main Survey. The entire SV data, internally known as Fuji, is publicly released as the DESI Early Data Release (EDR; [36]) while the First Data Release (DR1; DESI Collaboration et al. 2025) was released in March 2025. This work relies on the EDR data as the complementary studies of the VAC presented in [128]. The expansion to DR1 is left for future work. The DR1 already showcases the DESI potential by the cosmological results from the full-shape analysis ([38]).

The scale of DESI observations is supported by software pipelines and products, which include imaging from the DESI Legacy Imaging Surveys ([146, 41]), a fully automatic spectroscopic reduction pipeline ([61, 122]), followed by a template-fitting pipeline to derive classifications and redshifts for each targeted source (Redrock³; [3]; Bailey et al. in prep.), and for the special case of QSOs (Brodzeller et al. 2023). Redrock provides the redshift (Z), redshift uncertainty (ZERR), a redshift warning bitmask (ZWARN), and a spectral type (SPECTYPE) to every target based on the best fit. The photometry in $g$, $r$, $i$, and $z$ bands is estimated from the Legacy survey (DR9) images with the Tractor⁴ inference modeling code ([84]). Mid-infrared photometry is derived via forced photometry of the unWISE co-adds based on this optical model ([94]).

DESI EDR VAC of physical properties

The DESI EDR redshift catalog consists of 2 847 435 sources ([36]). For a sample of 1 337 250 galaxies and quasars (with SPECTYPE = GALAXY | QSO) characterized by reliable redshift (ZWARN = 0 or 4 and that do not have any fiber issues (COADDFIBERSTATUS = 0) we created a VAC of physical properties, including stellar masses and star formation rates (SFRs) as well as AGN features ([128]). Here, we present a short summary of the VAC and refer the reader to [128] for the full description.

Physical properties of DESI galaxies are derived by performing SED fitting using Code Investigating GALaxy Emission (CIGALE v2022.1; [11]) relying on the optical-MIR photometry and spectroscopic redshifts. CIGALE already proved its efficiency in deriving physical properties of galaxies hosting AGN (e.g., [118, 109] (Malek et al. 2018; Csizi et al. 2024)), as well as high-z AGN (e.g., [30, 143, 17, 101, 100]). CIGALE uses a library of models to fit the SED of galaxies, where the best fit is found based on the reduced $\chi^2$ minimization among all the possible combinations of the SED models. The library of models is created under the assumption of a delayed SFH with an optional exponential burst ([26]) and the Bruzual & Charlot [15] SSP models adopting a Chabrier [21] initial mass function. We assume solar metallicity and use Fritz et al. [52] models to account for possible AGN contribution. The grid of input parameters used to generate the model library follows the configuration adopted in the DESI EDR VAC (Siudek et al. [128]), and is summarized in Table A.1 in Appendix § 6. The generated models cover a wide range of objects, including galaxies without AGN contribution, as well as BL-, and NL-AGN. This provides flexibility and allows us to build the catalog of physical properties for both AGN and non-AGN host galaxies.

The SEDs of observed galaxies are created from the ground-based optical and near-infrared (NIR) photometry (i.e., $grz$ photometry, which we shortly refer to as optical) complemented by observations from MIR bands at 3.4, 4.6, 12 and 22 $\mu$m provided by the Wide-field Infrared Survey Explorer (WISE; Wright et al. [140]), a mission extension NEOWISE-Reactivation forced-photometry (Mainzer et al. [88]).

We note that the precise AGN classification of individual sources – especially those with low AGN fractions – as well as the stellar masses and SFRs may depend on the model assumptions, and the incorporation of the WISE photometry (see Siudek et al. [128] for a discussion). Based on the representative sample of $\sim50{,}000$ galaxies including passive, star-forming, and AGN galaxies, Siudek et al. [128] find that without WISE photometry the AGN fraction is overestimated (AGNFRAC $\geq0.1$) for 68% of star-forming galaxies selected based on the [NII]-BPT emission line diagnostic diagram (see details in Sect. 3.1). The requirement of a high SNR in the WISE observations (forced by FLAGINFRARED = 4) reduces it to 19%. This suggests that the MIR information is crucial to properly identify AGN via SED fitting. Siudek et al. [128] also show that the overestimation of the AGNFRAC for the star-forming galaxies without WISE information has a negligible effect on their stellar mass or SFR estimates. We further discuss the dependence of the AGN identification on the model choice in Appendix § 8.

Sample selection based on the AGN diagnostics

In this paper, we rely on the sample of 510 938 DESI EDR galaxies at $z \leq0.5$ from the VAC of physical properties⁵. We remove sources with poor photometric quality, defined as those with LOGM = 0, and sources with poor SED fits, defined as those with CHI2 $> 17$ (see Siudek et al. [128]). These criteria, along with the redshift limit (Z $\leq0.5$), define our parent sample. No additional cuts are applied at this stage; further selection criteria (e.g., SNR for emission-line quality or WISE-based selections) are introduced in the relevant sections.

To validate the performance of the SED-based AGN identification in recovering true AGN populations, we use a set of classical AGN diagnostics as reference. These include emission-line diagnostics (BPT-AGN), MIR colors (WISE-AGN), and high-energy diagnostics such as X-ray and radio emission. These methods are applied to galaxies at $z \leq0.5$, where line measurements and MIR photometry are most reliable. While the MIR color-based AGN selection draws from a subset of the photometric data available to full SED fitting, it remains a valuable cross-check. Its inclusion in this analysis serves primarily as a benchmark, given its common use and well-calibrated thresholds (e.g., Stern et al. [132]; Assef et al. [4]). We do not position it on equal footing with spectroscopy-based diagnostics or full SED fitting, as the latter incorporate broader wavelength coverage and physical modeling. Nonetheless, MIR color selection remains relevant for assessing completeness and consistency in AGN selection across surveys. To construct the BPT diagrams and BL-AGN sample, we use the emission line measurements from the DESI EDR FastSpecFit Spectral Synthesis and Emission-Line Catalog (FastSpecFit version 3.2 Fuji production⁶; Moustakas et al. [106]; Moustakas et al. in prep.). FastSpecFit is a stellar continuum and emission-line fitting code optimized to model jointly DESI optical spectra and broad-band photometry using physically motivated stellar continuum and emission-line templates⁷. The selection of the AGN sample is summarized in Sect. 3.5.

BPT diagram

The BPT diagrams (Baldwin et al. [5]) are widely used diagnostics relying on the strength of Balmer lines to forbidden nebular transitions to differentiate between star-forming galaxies and AGN. In this work, we use three diagrams based on the ratios of [NII]$\lambda6583$/H$\alpha$ vs. [OIII]$\lambda5007$/H$\beta$ ([NII]-BPT), [SII]$\lambda6717,6731$/H$\alpha$ vs. [OIII]$\lambda5007$/H$\beta$ ([SII]-BPT), and [OI]$\lambda6583$/H$\alpha$ vs. [OIII]$\lambda6300$/H$\beta$ ([OI]-BPT) similarly as Mezcua & Domínguez Sánchez [99]. The [NII]-BPT diagram is sensitive to metallicity, making AGN identification in metal-poor galaxies difficult (e.g., Groves et al. [57]; Cann et al. [20]). The [SII]-BPT offers better separation at intermediate metallicities but struggles with composite galaxies. The [OI]-BPT is more robust across metallicities but relies on a weak line that’s often hard to detect in faint or low-metallicity galaxies (Polimera et al. [112]). Recently, Ji & Yan 2020 proposed a simple reprojection of the [NII]-, [SII]-, and [OI]-BPTs that shows the potential to remove the ambiguity for the true composite objects. We leave the examination of this approach for future work.

The [NII]-, [SII]-, and [OI]-BPTs for the DESI galaxies are shown in Figure 1. We use the narrow components of the Balmer lines and apply a flux SNR $\geq3$ in each of the emission lines used for the respective [NII]-, [SII]-, and [OI]-BPT diagrams, except for [OI]$\lambda6300$, for which we use a flux SNR $\geq1$. The [OI]$\lambda6300$ line is weaker, and imposing a higher SNR threshold could exclude valid candidates (Cid Fernandes et al. [25]). The SNR cuts are applied separately for each diagram, meaning that the galaxies included in each BPT diagram may differ.

!Emission line diagnostic diagrams for DESI galaxies: N[II]-BPT (left), [SII]-BPT (middle), and [OI]-BPT (right) bbox=0.22,0.10,0.85,0.31

We also use the WHAN diagram introduced by Cid Fernandes et al. [25] that considers equivalent width (EW) of H$\alpha$ (EW(H$\alpha$)) vs. [NII]$\lambda6583$/H$\alpha$ line ratios. The WHAN diagram differentiates AGN (with EW(H$\alpha$) $\geq3$ and log[NII]/H$\alpha> -0.4$) from galaxies with hot old (post-AGB) stars that can mimic LINERs. To ensure reliable measurements, we require a flux SNR $\geq3$ for all relevant lines (H$\alpha$ and [NII]$\lambda6583$), as well as for EW(H$\alpha$). Consequently, the WHAN sample is not necessarily the same as the one used for the BPT diagrams, as the SNR cuts are applied independently.

BL-AGN selection

We identify BL-AGN at $z \leq0.5$ based on the presence of a broad H$\alpha$ component. Following the criteria outlined in Pucha et al. [115], we require SNR $\geq3$ for the total H$\alpha$ flux (SNR$_{\mathrm{HALPHA\_FLUX}}$), the broad H$\alpha$ component (SNR$_{\mathrm{HALPHA\_BROAD}}$), and the amplitude-over-noise (AoN) of the broad component (AoN$_{\mathrm{HALPHA\_BROAD}}$).

WISE selection

To select MIR-AGN we rely on the WISE color-color diagram $W2-W3$ vs. $W1-W2$ proposed by Hviding et al. [70]. We restrict the sample to galaxies with SNR $\ge3$ in all WISE bands used in the diagram and to $z \le0.5$. We note, that the forced WISE photometry can suffer from confusion and other systematic errors, which can impact the quality of the fitting and the resulting inference as well as the WISE SNR cuts. In addition, Legacy Surveys WISE photometry is quite shallow, especially in W3, and W4 and especially for emission line galaxies and other higher-redshift galaxies, which impacts the ability to identify AGN from the broadband SED modeling. The selection criteria proposed by Hviding et al. [70] are more sensitive to low luminosity and heavily obscured AGN than the traditional cuts given by Jarrett et al. [73] or Stern et al. [132]. However, WISE-AGN samples may be contaminated by star-forming galaxies diluting the WISE color-selected AGN [63].

X-ray and radio counterparts

We search for X-ray counterparts within 5 arcsec of the optical position of each source as given by the DESI Legacy Imaging Surveys making use of the Chandra Source Catalog (CSC DR2; Evans et al. [50, 49]), the XMM-Newton Serendipitous Source Catalogue (4XMM-DR13; Webb et al. [138]), and the all-sky survey catalog from the extended ROentgen Survey with an Imaging Telescope Array (eROSITA) on the Spectrum-Roentgen-Gamma (SRG) mission (eRASS1; Merloni et al. [97]). For identification of the radio counterparts within 5 arcsec from the sources we rely on the second data release of the LOFAR Two-metre Sky Survey (LoTSS DR2; Shimwell et al. [123]), the Faint Images of the Radio Sky at Twenty-cm (FIRST) Survey Final Catalog (Helfand et al. [67]) and the 1.4 GHz National Radio Astronomy Observatory (NRAO) Very Large Array (VLA) Sky Survey (NVSS; Condon et al. [29]).

AGN sample

We apply the optical and MIR selection criteria described above to classify 510 939 DESI EDR galaxies into four types. These four classes are used as a reference to estimate the best threshold on AGNFRAC for identifying AGN based on the SED fitting:

1. BPT-AGN: defined as AGN or composite in the [NII]-BPT, AGN on the WHAN diagram, and AGN in the [SII]- or [OI]-BPT. Selected number of sources: 6 019.

2. BPT-SF: defined as star-forming in the [NII]-, [SII]- and [OI]-BPTs. Selected number of sources: 122 731.

3. BL-AGN: defined as sources with a broad $H\alpha$ emission line. Selected number of sources: 4 065.

4. WISE-AGN: defined as AGN based on the WISE color diagram. Selected number of sources: 9 253.

Each class includes galaxies at $z \le0.5$ spanning a wide range of stellar masses. As shown in Figure 2, the BPT-SF population predominantly consists of lower-mass, lower star-formation galaxies at lower redshifts. Additionally, since SFRs are derived from CIGALE using AGN-inclusive templates, even BPT-SF galaxies may show slightly suppressed SFRs if the fit assigns any AGN contribution – this is especially relevant in ambiguous or composite systems. We explore also whether the $\chi^2$ distribution varies across different AGN and star-forming galaxy classifications (bottom right panel in Figure 2). Contrary to some earlier works (e.g., Berta et al. [8]) where elevated $\chi^2$ values – especially in stellar or star-forming-only fits – could signal the presence of a hidden AGN (e.g., visible in the NIR excess), we find that the distributions of reduced $\chi^2$ are remarkably similar for BPT-AGN, WISE-AGN, BL-AGN, and BPT-SF galaxies. This is likely due to the inclusion of AGN templates in the CIGALE fits, which improves the overall model match even in AGN-dominated sources. This suggests that $\chi^2$ is not a discriminant for AGN identification in our case, provided that proper AGN templates are used during the fitting process. Figure 3 and Table 1 highlight the diversity and overlap between AGN identified through different selection techniques: BPT-AGN, BL-AGN, and WISE-AGN. The significant areas of non-overlap (e.g., 66.8% of WISE-AGN are identified only by WISE colors) underscore that each method may be sensitive to different types of AGN and highlight the need for a multi-wavelength approach to identify the complete AGN sample as already investigated by e.g., Cann et al. [20] and Hviding et al. [70].

Classification	$N_{\mathrm{total}}$	$N_{\mathrm{unique}}$	$\%_{\mathrm{unique}}$	$N_{\mathrm{BPT-AGN}}$	$\%_{\mathrm{BPT-AGN}}$	$N_{\mathrm{BL-AGN}}$	$\%_{\mathrm{BL-AGN}}$	$N_{\mathrm{BPT-SF}}$	$\%_{\mathrm{BPT-SF}}$
BPT-AGN	6 019	3 438	57.1	–	–	1 835	30.5	0	0.0
BL-AGN	4 065	979	24.1	1 835	45.1	–	–	103	2.5
WISE-AGN	9 253	6 184	66.8	1 818	19.6	2 323	25.1	1 907	20.6

Results and Discussion

In this Section, we validate the AGN fraction (AGNFRAC) estimated via SED fitting of optical ($grz$) and MIR ($W1$4) observations as a proxy to identify AGN in comparison to standard techniques. As a large fraction of AGN is not detected in optical surveys due to either distance or dust obscuration (e.g., Truebenbach & Darling [137]), we validate how crucial it is to include the MIR photometry in SED fits to recreate the AGN contribution in galaxies. We also identify AGN candidates selected by the SED method that do not appear on traditional AGN diagrams.

SED fitting: AGN fraction

Usually, SED-based AGN selection is based on a fixed threshold on the AGN fraction derived from SED fitting (Thorne et al. [135]; Best et al. [9]; Bichang’a et al. [10]; Das et al. [34]). The AGN fraction is defined here as the fraction of the IR emission coming from the AGN to the total IR emission. We rely on the AGN and star-forming galaxy sample selected based on the traditional techniques (see Sect. 3.5) to find the best threshold for the AGN selection. Figure 4 shows the distribution of AGNFRAC for galaxies with different infrared coverage. We split the sample based on the FLUXINFRARED value, which indicates the number of WISE bands with high SNR ($\mathrm{SNR} \ge3$): a value of 4 means detections in all four bands, while lower values correspond to fewer reliable measurements. Although AGNFRAC uncertainties are computed by CIGALE, they are not included in the DESI EDR VAC and in this analysis because, in cases with faint AGN emission or limited infrared data, the error estimates can become poorly constrained and unreliable for robust statistical use. This limitation is inherent to SED fitting in such regimes and does not affect the broader AGN selection performance (see also Ciesla et al. 2015, Yang et al. 2020, and Mountrichas et al. 2021b). In the left panel (galaxies observed at most in two WISE bands with SNR $\ge3$, i.e., with FLAGINFRARED $\le2$), AGN and star-forming galaxies show nearly indistinguishable AGNFRAC distributions (typical values around 0.3), indicating unreliable AGN fraction estimates when only one or two WISE bands are available. In contrast, the right panel (galaxies observed in all four WISE bands with SNR $\ge3$, i.e., with FLAGINFRARED $= 4$) reveals clear separation: star-forming galaxies show low AGNFRAC values (median $= 0.02$), while AGN classes (BPT-AGN, BL-AGN, WISE-AGN) have broader distributions with higher medians (0.19, 0.12, and 0.25, respectively). Notably, BPT-AGN exhibit a bimodal distribution peaking at 0 and 0.1.

To further evaluate the discriminative power of AGNFRAC, we perform a classification analysis using Receiver Operating Characteristic (ROC; Bradley [12]) curves across our samples. The performance metrics and threshold analysis (see Appendix § 9) suggest that AGNFRAC is an effective AGN indicator only when applied to sources with reliable infrared measurements.

Taking into account the strong correlation between AGNFRAC and the availability of WISE photometry with SNR $\ge3$ (see also Sect. 4.2), we recommend applying a selection criterion of AGNFRAC $\ge0.1$ combined with FLAGINFRARED $= 4$ to define AGN based on SED fitting (SED-AGN). This threshold is motivated both by the observed peak in AGNFRAC distributions for spectroscopically confirmed AGN classes (see Figure 4) and by its optimal diagnostic performance in the ROC analysis (see Appendix § 9). For more inclusive samples, the infrared quality condition can be relaxed to FLAGINFRARED $\ge3$, which preserves the general distribution shape and yields similar median AGNFRAC values.

Importance of the MIR photometry

As shown in the previous Section, the distribution of AGNFRAC is sensitive to the availability of WISE photometry. We further verify the dependence of the BPT classification on the availability of the WISE photometry. Figure 5 shows the [NII]-BPT diagram for galaxies observed in all four WISE bands with SNR $\ge3$ (i.e., FLAGINFRARED $= 4$; right panel) and galaxies with at most two WISE observations with SNR $\ge3$ (i.e., FLAGINFRARED $\le2$; left panel). The distribution of AGNFRAC is almost uniform on the BPT diagram for galaxies without high SNR WISE photometry. Contrarily, for galaxies with FLAGINFRARED $= 4$, the AGNFRAC is gradually changing from $\lesssim0.1$ in the star-forming galaxies locus to $\gtrsim0.4$ in the region attributed to AGN.

Using the AGN fraction derived with CIGALE as the criterion to select AGN (AGNFRAC $\ge0.1$), we recover 74% of BPT-selected AGN (69% if limiting to sample with FLAGINFRARED $= 4$). However, at the same time, 62% star-forming galaxies also are identified as AGN (i.e., fulfilling the criterion AGNFRAC $\ge0.1$). The percentage of misclassified star-forming galaxies drops to 15% for a sample with FLAGINFRARED $= 4$. We note that Siudek et al. [128] show that the overestimation of the AGN fraction for misclassified star-forming galaxies (with AGNFRAC $\ge0.1$) does not affect the estimates of the stellar masses and SFRs (see Sect. 2.2).

Figure 6 shows the WISE diagram as a function of the AGNFRAC depending on the availability of the WISE photometry. There is no visual difference in the dependence on the AGNFRAC whether the sample is limited to SNR $\ge3$ in W1–3; i.e., FLAGINFRARED $= 3$ or all WISE bands (W1–4; i.e., FLAGINFRARED $= 4$), but the sample size differs significantly.

The sample selected with FLAGINFRARED ≥ 3 includes 9 253 sources, while restricting to FLAGINFRARED = 4 results in 4 974 galaxies. Without any restriction on the FLAGINFRARED, the WISE-AGN are recovered well with SED selection (79% of WISE-AGN are identified as SED-AGN). At the same time, 40% of non-AGN are identified as SED-AGN. Only when applying a more strict cut on WISE photometry (FLAGINFRARED = 4) leads to the identification of 86% of WISE-AGN and reduces the misclassified AGN to 33%.

We find a similar correlation of the AGNFRAC with WISE colors as found for the BPT selection: the region attributed to AGN is characterized by a high AGN fraction (with a median AGNFRAC = 0.25). Outside the AGN selection region, the AGN fraction is low (AGNFRAC ≲ 0.07). However, there are two regions attributed to higher AGN fraction (AGNFRAC ∼ 0.5) on the right-low end and left-upper of the AGN envelope. While the upper-left region is sparsely populated and can be neglected, the tail in $W2-W3$ ($W2-W3 \lesssim2.5$) across $0.5 \lesssim W1-W2 \lesssim0$ color gathers 2 017 galaxies (i.e., half of the WISE-AGN sample; SED-MIR-AGN candidates hereafter).

SED-MIR-AGN candidates are in the majority (91%) characterized by high AGN fraction (AGNFRAC ≳ 0.5)⁸. They are massive passive galaxies (with median $\log(M_{\mathrm{star}}/M_{\odot}) \sim11$ and median $\log(\mathrm{SFR}/M_{\odot}\,\mathrm{yr}^{-1}) \sim-6$). SED-MIR-AGN candidates show tentative⁹ signatures of [NeIII]λ3869 emission (with median full width at half maximum similar to the one found for AGN, i.e., FWHM ∼ 320 km s$^{-1}$, while for non-AGN FWHM is ∼ 230 km s$^{-1}$). The [NeIII]λ3869 emission together with [OIII]λ3727, 3729 emission line and $g-z$ rest-framed color forms another AGN diagnostic diagram (TBT diagram; Trouille et al. [136]). Trouille et al. [136] showed that the TBT diagram is efficient in recreating [NII]-BPT selection (98.7% of BPT-AGN are identified as TBT-AGN and 97% of the BPT-SF as TBT-SF) at the same time outperforming the BPT in identifying X-ray sources, while restricting to FLAGINFRARED = 4 results in 4 974 galaxies. Without any restriction on the FLAGINFRARED, the WISE-AGN are recovered well with SED selection (79% of WISE-AGN are identified as SED-AGN). At the same time, 40% of non-AGN are identified as SED-AGN. Only when applying a more strict cut on WISE photometry (FLAGINFRARED = 4) leads to the identification of 86% of WISE-AGN and reduces the misclassified AGN to 33%.

at the same time outperforming the BPT in identifying X-ray selected AGN (TBT identifies 97% of the X-ray AGN as TBT-AGN, while BPT only 80%). Almost all SED-MIR-AGN candidates (94%) are identified as AGN according to the TBT diagram, however, we cannot be confident of their AGN nature, as only 65 have [NeIII]λ3869 detected with SNR ≥ 3. The weak and noisy [NeIII]λ3869 may bias the conclusion.

SED-AGN sample

Table 2 shows that the SED-AGN sample, defined by galaxies with AGNFRAC ≥ 0.1 and FLAGINFRARED = 4 at $z \le0.5$, identifies a significant number of AGN (≥ 70%) selected using traditional methods. This overlap suggests that the SED-fitting approach is capable of capturing a broad range of AGN types. As detailed in Table 2, the SED-based classification successfully identifies around 70% of BL-AGN and BPT-AGN and over 85% of WISE-AGN. This high level of correspondence demonstrates the robustness of the SED method in capturing classical AGN types. However, about 15% of star-forming galaxies (BPT-SF) are also classified as AGN by the SED method, indicating some level of contamination from non-AGN sources.

Category	N	SED-AGN Matches %	SED-AGN Only %
BL-AGN	2 253	70.13
BPT-AGN	2 703	69.29
WISE-AGN	4 953	86.45
BPT-SF	20 717	15.25
SED-AGN	25 123	-	51.78

A particularly notable result is that over half of the SED-AGN sample (∼ 52%) is not identified by any of the main AGN selection methods (BL-AGN, BPT-AGN, WISE-AGN, or BPT-SF). This raises an important question: is the SED-based method producing a high rate of false positives, or is it effectively uncovering a population of AGN that are missed by traditional diagnostics? To better understand the nature of the SED-AGN sample, we expanded our analysis beyond the standard AGN selection techniques. Specifically, we incorporate additional AGN classifications derived from all three BPTs, X-ray, and radio AGN diagnostics. This broader approach allows for a more comprehensive assessment of the AGN content within the SED-selected population.

When incorporating these additional AGN identification methods, the fraction of galaxies classified solely by the SED criteria (SED-AGN Only) decreases significantly – from 51.78% to 16.14% (see Table 3), highlighting both the complementary value and the limitations of the SED-based method. The SED-based approach appears to at least partially recover AGN populations that are overlooked by traditional diagnostics, emphasizing the value of SED fitting in building a more complete and inclusive AGN census. However, at the same time, a non-negligible fraction of the SED-AGN sample ($\sim15$%) overlaps with star-forming galaxies (BPT-SF), suggesting some level of contamination. Moreover, the AGN nature of sources classified solely as SED-AGN remains unconfirmed (see discussion below), challenging the use of SED fitting for constructing clean and pure AGN samples.

Category	N	SED-AGN Matches %	SED-AGN Only %
NII-AGN	3 791	67.95
NII-SF	23 175	17.48
NII-LINER	1 064	54.42
NII-COMPOSITE	7 939	33.54
SII-AGN	2 583	68.45
SII-SF	31 458	21.13
SII-LINER	630	53.65
OI-AGN	3 510	69.66
OI-SF	30 420	22.72
OI-LINER	832	62.26
WHAN-AGN	32 283	38.61
WHAN-SF	25 534	20.36
WHAN-RG	3 559	67.04
X-RAY-AGN	793	45.27
RADIO-AGN	5 326	38.64
SED-AGN	4 054	-	16.14

However, we note that the nature of LINERs remains debated, with alternative ionization sources such as post-asymptotic giant branch (post-AGB) stars or shocks being proposed in the literature (e.g. Stasińska et al. 2008; Singh et al. 2013; Belfiore et al. 2016).

The SED-based technique identifies roughly 50% of LINERs as selected by the BPT diagrams. The interpretation of LINERs remains actively debated in the literature, with studies supporting both AGN (e.g., [65, 78]) and stellar photoionization scenarios (e.g., [131, 120]). The overlap with LINERs suggests that the SED-AGN method is sensitive to a diverse AGN population, including those in different evolutionary stages or in obscured environments. The SED-based method maintains a relatively low level of contamination except in the WHAN diagram. The SED-AGN identifies 40% of WHAN-AGN but also 67% of retired galaxies. Interestingly, the majority (96%) of these retired SED-AGN are actually "liny" (i.e., showing the presence of the emission lines; 0.5 $\leq$ EW(H$\alpha$) $< 3$ Å) galaxies. The WHAN diagram is known to struggle with distinguishing low-luminosity AGN from "liny" retired galaxies, where line emission could originate from past AGN activity or low-level star formation ([25, 24, 68]).

This raises the possibility that the SED-AGN method may be identifying galaxies transitioning from AGN activity to a more passive state, thus appearing as retired on the WHAN diagram. Supporting this interpretation, Herpich et al. [68] suggested that "liny" retired galaxies might have undergone recent star formation, and their emission line gas could originate from the AGN activity. Similarly, Agostino et al. [1] show that nearly one-third of X-ray AGN fall in the WHAN "retired" locus. In our sample, out of 2386 SED-AGN classified as retired in WHAN, 515 (22%) are independently confirmed as AGN via other diagnostics, further questioning their classification as fully passive systems.

The X-ray data (see Sect. 3.4) reveals 882 sources with X-ray emission. Using the star formation rates derived from CIGALE SED fitting and the relation from Lehmer et al. [85], we estimate the expected 2–10 keV luminosity from star formation and find that 793 galaxies exceed this by more than 3$\sigma$, confirming AGN-like X-ray emission (see also Mezcua et al. [98]). The SED-based method identifies $\sim50$% of these X-ray AGN, even though X-ray data are not used in the fitting, demonstrating good sensitivity to AGN across different regimes. To understand why the other 50% are missed, we examine their classification using other methods. Most are also missed by traditional BPT, and WISE diagnostics, with nearly 40% falling into the WHAN-AGN region. This suggests that most missed X-ray AGN are generally low-luminosity or obscured, escaping both SED and classical diagnostics, and reinforces the known difficulty of identifying such AGN without deep X-ray data. This is consistent with previous studies such as Trouille et al. [136] and Juneau et al. [75, 76], which showed that only $\sim20$–30% of X-ray AGN fall in the AGN region of the BPT diagram, with many appearing as star-forming or composite systems instead. These studies concluded that emission-line diagnostics become increasingly incomplete for identifying obscured or radiatively inefficient AGN. Our results are in agreement with these findings, and show that SED fitting adds significant value by recovering $\sim50$% of X-ray AGN, highlighting its effectiveness in such regimes.

Of 13 450 galaxies with radio detections (see Sect. 3.4), 5 326 exhibit radio luminosities either above $5 \times10^{23}$ W/Hz or more than $3\sigma$ above what is expected from star formation (e.g., Mezcua et al. [102]), classifying them as radio AGN. The SED fitting recovers 40% of these sources. Similar to the X-ray case, the radio AGN missed by SED tend to be weakly active or obscured, with most showing no overlap with BPT, or WISE diagnostics, and $\sim45\%$ falling within the WHAN-AGN region. This again highlights the complementary but incomplete nature of the SED-based classification and reinforces the need for multi-wavelength data to uncover the full AGN population. Moreover, the SED fitting depends on model assumptions, which affect the AGN classification (see Table C.1 in Appendix § 8 for details).

Figure 7 shows the stacked spectra of BPT-AGN, WISE-AGN, SED-AGN, and BPT-SF galaxies. The spectra were stacked using mean normalization, preserving the relative fluxes of the emission lines, which is essential for a reliable comparison of spectral features across different samples. The SED-AGN stack shows similarity to the BPT-AGN and WISE-AGN stacks, particularly evident in the characteristic AGN emission line ratios, such as [NII]/H$\alpha$. Unlike the BPT-SF sample, where the [NeV]$\lambda$3427 emission is absent, the presence and relative strength of [NeV]$\lambda$3427 in SED-AGN align more closely with the AGN-dominated spectra. High-ionization lines, such as [NeV]$\lambda$3427, require energies well above the limit of stellar emission (55 eV; e.g., Leitherer et al. [86]; Izotov et al. [72]) and can be an efficient tracer of the AGN (e.g Gilli et al. [56]; Negus et al. [107]; Barchiesi et al. [6]).

SED-AGN Only: properties of the host galaxies

Despite the effectiveness of the SED-based technique, 4 054 out of 25 123 SED-AGN galaxies (16.14%, see Table 3) are classified as SED-AGN Only, meaning they are not confirmed by any other AGN selection method. To further explore the nature of these SED-AGN Only galaxies, we analyze their properties in comparison to properties of the BPT-AGN, WISE-AGN, SED-AGN, and BPT-SF galaxies¹⁰ as summarized in Table 4. These SED-AGN Only galaxies tend to be more massive, with median stellar masses comparable to those of BPT-AGN and WISE-AGN hosts ($\log(M_{\star}/M_{\odot}) \sim10.75$). However, they exhibit significantly lower star formation rates ($\log(\mathrm{SFR}/M_{\odot}\mathrm{yr}^{-1}) \sim-2$) than BPT-AGN and WISE-AGN, suggesting that they are generally passive galaxies.

Unit	BPT-SF	BPT-AGN	WISE-AGN	SED-AGN	SED-AGN Only
N	20 717	2 703	4 953	25 123	4 054
Z	$0.19^{+0.09}_{-0.08}$	$0.28^{+0.09}_{-0.09}$	$0.29^{+0.09}_{-0.06}$	$0.31^{+0.09}_{-0.08}$	$0.29^{+0.12}_{-0.09}$
LOGM [$\mathrm{M}_{\odot}$]	$9.82^{+0.33}_{-0.37}$	$10.61^{+0.25}_{-0.28}$	$10.48^{+0.32}_{-0.38}$	$10.58^{+0.25}_{-0.29}$	$10.75^{+0.26}_{-0.31}$
LOGSFR [$\mathrm{M}_{\odot}\mathrm{yr}^{-1}$]	$0.87^{+0.37}_{-0.40}$	$1.10^{+0.42}_{-0.53}$	$1.14^{+0.41}_{-0.46}$	$0.69^{+0.52}_{-1.31}$	$-2.23^{+2.10}_{-3.33}$
D4000	$1.22^{+0.06}_{-0.05}$	$1.33^{+0.17}_{-0.20}$	$1.15^{+0.14}_{-0.12}$	$1.22^{+0.14}_{-0.11}$	$1.76^{+0.23}_{-0.32}$

To confirm the passive nature of SED-AGN Only hosts, we also check the 4000 Å break (D4000), an indicator of the age of the stellar population: a high D4000 value (D4000 $> 1.5$) indicates older stars and high metallicity, while a lower value (D4000 $\le1.5$) suggests younger, star-forming populations (e.g., Kauffmann et al. [77]; Siudek et al. [127, 126]). The high D4000 values of the SED-AGN Only (D4000 = 1.76) indicate that old stars dominate the light from these galaxies. Moreover, we find that the majority (78%) of the AGN-SED Only show red colors on the $U$-$V$ vs. $V$-$J$ diagram (Whitaker et al. [139]; Siudek et al. [128]), consistent with their passive nature. This contrasts sharply with other classes, where the fraction of red galaxies is much lower, ranging from 0% in star-forming galaxies to 13% in BPT-AGN. Also, the optical part of the SEDs of SED-AGN Only galaxies is dominated by old stellar populations, while AGN contribution is seen in the MIR (see Appendix § 7). This analysis suggests that SED-AGN Only galaxies may represent a population of galaxies transitioning away from active star formation, possibly hosting weak or fading AGN activity that other methods fail to detect.

We also verify whether there are signatures of non-AGN nature among SED-AGN Only. Only 15 of SED-AGN Only ($< 1\%$) are characterized by line ratios characteristic of shock excitation, such as [SII]$\lambda$6717+$\lambda$6731/H$\alpha> 0.4$ or [OI]/H$\alpha> 0.1$, which are traditionally associated with supernova remnants or other shock-heated regions rather than photoionization by an AGN (e.g., Dodorico et al. [42]; Rich et al. [117]; Comerford et al. [28]). This suggests that contamination from non-AGN processes in the SED-AGN Only sample is minimal.

Summary

We identify AGN based on the AGN fraction (AGNFRAC $\ge0.1$) derived from SED fitting with CIGALE applied to DESI EDR galaxies at $z \le0.5$. As discussed in Sect. 1, traditional AGN identification diagnostics, such as the BPT diagram, are limited by luminosity biases, often leading to incomplete samples. In contrast, the SED-based approach, leveraging a multiwavelength strategy, promises a more comprehensive identification of AGN. Our key findings are summarized as follows:

• Importance of WISE photometry: MIR photometry is essential for robust AGN identification via AGNFRAC. When high SNR (SNR $\ge3$) WISE photometry is limited (i.e., FLAGINFRARED $\le2$), 62% of BPT-selected star-forming galaxies are incorrectly classified as AGN. Including all four WISE bands with SNR $\ge3$ (FLAGINFRARED = 4) reduces this contamination to 15%. Meanwhile, 70% of BPT-AGN are consistently recovered, irrespective of the WISE photometry’s SNR (see Sect. 4.2).

• MIR-AGN candidates beyond WISE diagram:: The SED fitting method identifies a distinct subset of strong AGN candidates (SED-MIR-AGN) that fall outside the canonical AGN region on the WISE diagram. These sources are confirmed as AGN in 98% of cases via the TBT diagram, although weak and noisy [NeIII]$\lambda$3869 emission limits definitive confirmation (see Sect. 4.2).

• Comparison with standard AGN selection methods: Our SED-based AGN classification shows substantial agreement with traditional techniques, recovering approximately 70% of both BL-AGN and BPT-AGN, and over 85% of WISE-selected AGN. However, the method does have a 15% contamination rate from star-forming galaxies.

• SED fitting as a unifying AGN diagnostic: The SED approach uncovers a population of AGN (SED-AGN Only) missed by standard diagnostics. Relying solely on BPT-AGN, WISE-AGN, and BPT-SF selections fails to identify $\sim52\%$ of SED-AGN (see Table 2). Incorporating multiple selection methods reduces this fraction to $\sim16\%$ (see Table 3). While promising, the SED-AGN Only population

includes uncertain cases (e.g., LINERs, “liny” retired galaxies), and may still be affected by a ∼15% contamination from star-forming galaxies.

– Dependence on model assumptions: The efficiency and purity of SED-based AGN selection are moderately sensitive to modeling choices (see Table C.1 and Fig. D.1). Changes to dust attenuation laws and AGN templates can alter the balance between completeness and contamination, while variations in IMF, SFH, and SSP models have smaller effects. We note that allowing metallicity to vary as a free parameter achieves similar or even slightly better performance, further reducing contamination from star-forming galaxies and marginally improving AGN selection efficiency. The use of mid-infrared data is crucial—removing WISE W3 and W4 bands leads to severe contamination by star-forming galaxies.

This work demonstrates the importance of multi-wavelength data fusion for robust AGN identification, with direct implications for upcoming large-scale surveys. The Euclid mission will provide deep near-infrared photometry across 14 000 deg$^{2}$ ([46]), LSST will deliver unprecedented optical depth and time-domain coverage ([71]), and the Spectro-Photometer for the History of the Universe, Epoch of Reionization, and ices Explorer (SPHEREx) will perform an all-sky spectroscopic survey with near-immediate public data release ([43, 31]). However, mismatched depths and sky coverage among these facilities present challenges for optimal data fusion (e.g., [95, 69]).

Our 70–80% recovery rate using DESI Legacy Surveys and WISE photometry demonstrates that systematic multi-wavelength approaches can effectively overcome individual survey limitations. To further enhance AGN selection, machine-learning techniques are proving to be increasingly powerful. Semi-supervised clustering applied to medium-resolution spectroscopy can automatically separate narrow-line and broad-line AGN with high accuracy ([126, 124, 45]). Autoencoder architectures successfully recover AGN candidates missed by traditional BPT diagnostics due to low-SNR line measurements (Alcolea et al., in prep.), while diffusion-based models identify AGN from single-band morphological analysis of Euclid optical images ([48]). Foundation models may further unify multi-wavelength AGN detection, though their full potential remains to be explored ([47]). The multi-wavelength approach, such as the systematic incorporation of SPHEREx spectroscopic data into LSST AGN selection pipelines, analogous to the DESI Legacy Surveys and WISE approach presented here, could reduce the 15% star-forming galaxy contamination observed in optical-only classifications. More broadly, a multi-modal approach that combines morphological, spectral, and photometric data may be key to robust AGN identification in the era of next-generation surveys.

Acknowledgements. The authors thank the anonymous referee for insightful comments. This work has been supported by the Polish National Agency for Academic Exchange (Bekker grant BPN/BEK/2021/1/00298/DEC/1) and the State Research Agency of the Spanish Ministry of Science and Innovation (grant PGC2018-100852-A-I00 and PID2021-126838NB-I00). M.M. acknowledges support from the Spanish Ministry of Science and Innovation through the project PID2021-124243NB-C22, and the program Unidad de Excelencia María de Maeztu CEX2020-001058-M. H.Z. acknowledges the support from the National Natural Science Foundation of China (NSFC; grant Nos. 12120101003 and 12373010) and National Key R&D Program of China (grant Nos. 2023YFA1607800, 2022YFA1602902) and Strategic Priority Research Program of the Chinese Academy of Science (Grant Nos. XDB0550100).

This material is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, Office of High-Energy Physics, under Contract No. DE–AC02–05CH11231, and by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility under the same contract. Additional support for DESI was provided by the U.S. National Science Foundation (NSF), Division of Astronomical Sciences under Contract No. AST-0950945 to the NSF’s National Optical-Infrared Astronomy Research Laboratory; the Science and Technology Facilities Council of the United Kingdom; the Gordon and Betty Moore Foundation; the Heising-Simons Foundation; the French Alternative Energies and Atomic Energy Commission (CEA); the National Council of Humanities, Science and Technology of Mexico (CONAHCYT); the Ministry of Science, Innovation and Universities of Spain (MICIU/AEI/10.13039/501100011033), and by the DESI Member Institutions: {link}https://www.desi.lbl.gov/collaborating-institutions. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the U. S. National Science Foundation, the U. S. Department of Energy, or any of the listed funding agencies. The authors are honored to be permitted to conduct scientific research on Iolkam Du’ag (Kitt Peak), a mountain with particular significance to the Tohono O’odham Nation. The DESI Legacy Imaging Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS), the Beijing-Arizona Sky Survey (BASS), and the Mayall z-band Legacy Survey (MzLS). DECaLS, BASS and MzLS together include data obtained, respectively, at the Blanco telescope, Cerro Tololo Inter-American Observatory, NSF’s NOIRLab; the Bok telescope, Steward Observatory, University of Arizona; and the Mayall telescope, Kitt Peak National Observatory, NOIRLab. NOIRLab is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation. Pipeline processing and analyses of the data were supported by NOIRLab and the Lawrence Berkeley National Laboratory. Legacy Surveys also uses data products from the Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE), a project of the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration. Legacy Surveys was supported by the Director, Office of Science, Office of High Energy Physics of the U.S. Department of Energy; the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility; the National Science Foundation, Division of Astronomical Sciences; the National Astronomical Observatories of China, the Chinese Academy of Sciences and the Chinese National Natural Science Foundation. LBNL is managed by the Regents of the University of California under contract to the U.S. Department of Energy. The complete acknowledgments can be found at {link}https://www.legacysurvey.org/.

Data Availability

The VAC of physical properties of DESI EDR galaxies is publicly available at {link}https://data.desi.lbl.gov/doc/releases/edr/vac/cigale/. The presented analysis is conducted on v1.4 based on the redshift coming from the QSO afterburner pipeline relying on the QuasarNet and the broad Mg II finder pipelines. The data behind the figures are available at {link}https://doi.org/10.5281/zenodo.15622223.

1 Instituto de Astrofísica de Canarias (IAC); Departamento de Astrofísica, Universidad de La Laguna (ULL), 38200, La Laguna, Tenerife, Spain

2 Institute of Space Sciences (ICE, CSIC), Campus UAB, Carrer de Can Magrans, s/n, 08193 Barcelona, Spain

3 Institut d’Estudis Espacials de Catalunya (IEEC), c/ Esteve Terradas 1, Edifici RDIT, Campus PMT-UPC, 08860 Castelldefels, Spain

4 Department of Physics & Astronomy, University College London, Gower Street, London, WC1E 6BT, UK

5 Institute of Cosmology and Gravitation, University of Portsmouth, Dennis Sciama Building, Portsmouth, PO1 3FX, UK

6 Department of Physics and Astronomy, Siena College, 515 Loudon Road, Loudonville, NY 12211, USA

7 National Astronomical Observatories, Chinese Academy of Sciences, A20 Datun Road, Chaoyang District, Beijing, 100101, P. R. China

8 Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, CA 94720, USA

9 Department of Physics, Boston University, 590 Commonwealth Avenue, Boston, MA 02215 USA

10 Dipartimento di Fisica “Aldo Pontremoli”, Università degli Studi di Milano, Via Celoria 16, I-20133 Milano, Italy

11 INAF-Osservatorio Astronomico di Brera, Via Brera 28, 20122 Milano, Italy

12 Department of Physics and Astronomy, The University of Utah, 115 South 1400 East, Salt Lake City, UT 84112, USA

13 Instituto de Física, Universidad Nacional Autónoma de México, Circuito de la Investigación Científica, Ciudad Universitaria, Cd. de México C. P. 04510, México

14 NSF NOIRLab, 950 N. Cherry Ave., Tucson, AZ 85719, USA

15 Departamento de Física, Universidad de los Andes, Cra. 1 No. 18A-10, Edificio Ip, CP 111711, Bogotá, Colombia

16 Observatorio Astronómico, Universidad de los Andes, Cra. 1 No. 18A-10, Edificio H, CP 111711 Bogotá, Colombia

17 Fermi National Accelerator Laboratory, PO Box 500, Batavia, IL 60510, USA

18 Department of Physics, The University of Texas at Dallas, 800 W. Campbell Rd., Richardson, TX 75080, USA

19 Department of Physics and Astronomy, University of California, Irvine, 92697, USA

20 Sorbonne Université, CNRS/IN2P3, Laboratoire de Physique Nucléaire et de Hautes Energies (LPNHE), FR-75005 Paris, France

21 Institució Catalana de Recerca i Estudis Avançats, Passeig de Lluís Companys, 23, 08010 Barcelona, Spain

22 Institut de Física d’Altes Energies (IFAE), The Barcelona Institute of Science and Technology, Edifici Cn, Campus UAB, 08193, Bellaterra (Barcelona), Spain

23 Instituto de Astrofísica de Andalucía (CSIC), Glorieta de la Astronomía, s/n, E-18008 Granada, Spain

24 Departament de Física, EEBE, Universitat Politècnica de Catalunya, c/Eduard Maristany 10, 08930 Barcelona, Spain

25 Department of Physics and Astronomy, Sejong University, 209 Neungdong-ro, Gwangjin-gu, Seoul 05006, Republic of Korea

26 CIEMAT, Avenida Complutense 40, E-28040 Madrid, Spain

27 Department of Physics, University of Michigan, 450 Church Street, Ann Arbor, MI 48109, USA

28 University of Michigan, 500 S. State Street, Ann Arbor, MI 48109, USA

29 Department of Physics & Astronomy, Ohio University, 139 University Terrace, Athens, OH 45701, USA

SED Fitting Parameters

To ensure transparency and reproducibility, we provide in Table A.1 the full set of input parameters and grid values used in the SED fitting process with CIGALE. These parameters follow the configuration adopted in the DESI EDR VAC (Siudek et al. [128]). As CIGALE-based analyses are sensitive to the choice and priors of model parameters, we use a grid that balances physical coverage with computational feasibility.

The SEDs for representative examples

Figure B.1 shows the SEDs for representative examples from each category: a typical star-forming galaxy, a WISE/BPT-AGN, and a SED-AGN Only galaxy. The star-forming galaxy exhibits a strong stellar component, with the SED peaking in the optical and NIR regions. The dust emission, indicated by the infrared excess, reflects ongoing star formation. The AGN contribution is negligible, consistent with the expected characteristics of a purely star-forming galaxy. The WISE/BPT-AGN, in contrast, shows significant AGN activity, particularly in the MIR range. The SED reveals a combination of strong stellar emission and enhanced IR output due to AGN-heated dust. This suggests that while the galaxy is actively forming stars, the AGN is also contributing significantly to the overall energy output. The SED-AGN Only galaxy presents a distinct profile, with a relatively weaker stellar component and a dominant AGN contribution in the MIR. The optical part of the SED suggests a lower star formation rate or an older stellar population. The strong MIR excess is primarily driven by the AGN, indicating that these galaxies are likely passive, where AGN activity dominates. These SED comparisons highlight the unique nature of SED-AGN Only galaxies, possibly representing a population of passive galaxies where the AGN is the dominant source of energy.

Dependence on model assumptions

When constructing the model grid for the SED fitting (see Sect. 2.2), several assumptions are made regarding the SFH, metallicity, and dust attenuation laws. Here, we test the robustness of AGN identification against changes in these model assumptions using a representative sample of approximately 50 000 galaxies from DESI EDR, covering a broad range of galaxy types (see Appendix D1 in Siudek et al. [128] for a description of the representative sample).

Table C.1 summarizes the efficiency of AGN identification under different model choices relative to the default configuration. We always change only one parameter at a time, keeping all others fixed (e.g., when applying a non-parametric SFH of Ciesla et al. [27], we retain the IMF from Chabrier [21] and SSP models from Bruzual & Charlot [15]). For a full description of our model components, see Sect. 6 in Siudek et al. [128]. We find the following:

• Changes in the IMF to Salpeter [119] or the dust emission model to Dale et al. [33] have negligible impact on the AGN selection efficiency.

• Adopting the Charlot & Fall [22] dust attenuation model reduces contamination by star-forming galaxies, but also decreases completeness for BPT-AGN and WISE-AGN.

• Using a non-parametric SFH (Ciesla et al. [27]) or the SSP models from Maraston [91] slightly increases the AGN identification rate, but at the cost of slightly increased contamination from star-forming galaxies.

• Allowing metallicity to vary as a free parameter improves both the completeness of AGN selection and reduces contamination.

• Switching the AGN emission model to SKIRTOR (Stalevski et al. [129, 130]) enhances completeness, notably for broad-line AGN (up to 92% recovery) and WISE-AGN (97%), but at the cost of increased contamination from star-forming galaxies.

The choice of model ultimately depends on the scientific goals. Increased contamination from star-forming galaxies is generally undesirable in AGN-focused studies, as it reduces sample purity and can bias physical interpretations. However, any model-driven enhancement or suppression of certain populations should be interpreted cautiously, as it may not reflect the true galaxy demographics. Thus, model selection should reflect the intended use—whether to maximize AGN completeness (e.g., for demographic studies) or to ensure purity (e.g., for SED-based analyses). In the next Section, we further quantify the performance of the AGN identification depending on the model assumptions.

We additionally note that when WISE W3 and W4 bands are excluded from the SED fitting, the contamination from star-forming galaxies rises significantly, with recovery rates dropping for BPT-AGN, WISE-AGN, and broad-line AGN. In extreme cases where no MIR information is used, virtually all galaxies are classified as AGN (AGNFRAC $\geq0.1$), illustrating the critical importance of MIR data in controlling contamination.

AGN selection using AGNFRAC

We assess the performance of the AGNFRAC parameter, derived from SED fitting, as a tool for AGN classification by performing a ROC analysis for: i) sources with high-quality mid-IR photometry (FLAGINFRภARED = 4), and ii) those with poor or missing WISE coverage (FLAGINFRARED $\leq2$). A ROC curve is a representation of the true positive rate (TPR) versus the false positive rate (FPR) for different classification thresholds. The TPR and FPR are defined as:

where TP, FP, FN, and TN are the counts of true positives (AGN that is classified as AGN), false positives (non-AGN that is classified as AGN), false negatives (AGN that is not classified as AGN), and true negatives (non-AGN that is not classified as AGN), respectively. A robust classification is achieved when AGN are classified as AGN (TPR = 1) rather than AGN as non-AGN (FNR = 0).

As shown in Fig. D.1, the ROC curve for the FLAGINFRARED = 4 subset yields a significantly higher area under the curve (AUC = 0.87¹¹), indicating strong separability between AGN and star-forming galaxies using AGNFRAC. In contrast, the ROC for the lower-quality subset (FLAGINFRARED $\leq2$) results in a much flatter curve with AUC = 0.57, close to random classification. Annotated thresholds on the ROC curves demonstrate that for FLAGINFRARED = 4, completeness increases sharply with minimal contamination, especially around AGNFRAC values of 0.1 – 0.3. In the lower-quality sample, no such threshold yields both high true positive and low false positive rates.

Parameter	Symbol	Values
Stellar population models: Bruzual & Charlot [15]
Initial mass function	IMF	Chabrier [21]
Metallicity	$Z$	0.02
SFH: Double exponentially decreasing
$\tau$ of the main stellar population (Gyr)	$\tau_{\mathrm{main}}$	0.1, 0.5, 1, 3, 5, 8
Age of the main stellar population (Gyr)	$t_1$	0.5, 1, 3, 4.5, 6, 8, 10, 13
$\tau$ of the burst stellar population (Gyr)	$\tau_{\mathrm{burst}}$	10
Age of the burst stellar population (Gyr)	$t_{\mathrm{burst}}$	0.05
Mass fraction of young stellar population	$f_{\mathrm{ySP}}$	0, 0.01, 0.1, 0.2
Nebular emission
Ionization parameter	$\log U$	-2
Gas metallicity	$Z_{\mathrm{gas}}$	0.02
Dust attenuation: Calzetti et al. [19]
Color excess of the nebular emission	$E(B-V)_{\mathrm{line}}$	0, 0.05, 0.15, 0.3, 0.5, 0.75, 0.9, 1.1, 1.3, 1.6
Reduction factor to apply on $E(B-V)_{\mathrm{line}}$	$E(B-V)_{\mathrm{star}}/E(B-V)_{\mathrm{line}}$	0.44
Dust emission: Draine et al. [44]
Mass fraction of PAHs	$q_{\mathrm{PAH}}$	0.47, 1.12, 2.5, 3.19
Minimum radiation field	$U_{\min}$	15
Power law slope of the radiation field	$\alpha$	2.0
Fraction illuminated from $U_{\min}$ to $U_{\max}$	$\gamma$	0.02
AGN: Fritz et al. [52]
The angle between the equatorial axis and line-of-sight	AGNPSY [deg]	0.001, 20.100, 40.1, 70.100, 89.990
Contribution of the AGN to the total LIR	AGNFRAC	0, 0.01, 0.1, 0.3, 0.5, 0.7, 0.9

Category	N	Default	IMF	Dust atten.	Dust emission	SFH	Z	SSP model	AGN model
BL AGN	1 033	78	79	75	75	80	83	81	92
BPT-AGN	1 479	65	66	55	65	70	67	75	80
BPT-SF	1 901	21	22	15	27	24	16	30	27
WISE-AGN	2 244	85	86	75	81	89	89	87	97
NII-AGN	2 215	64	66	53	65	70	66	75	78
NII-SF	2 118	26	26	19	30	29	20	34	32
NII-LINER	655	47	47	43	52	48	50	62	56
NII-COMPOSITE	3 543	29	30	20	36	33	26	43	40
SII-AGN	1 473	64	65	55	66	69	67	75	79
SII-SF	6 336	30	31	21	36	34	27	42	40
SII-LINER	360	44	44	48	46	45	52	62	54
OI-AGN	1 700	65	66	56	65	70	68	75	80
OI-SF	6 311	31	31	21	37	34	27	42	40
OI-LINER	333	48	48	50	46	48	52	63	57
WHAN-AGN	6 301	40	40	29	45	44	38	52	51
WHAN-SF	2 069	34	34	28	35	36	30	41	42
WHAN-RG	876	51	51	50	49	50	53	63	59
X-RAY-AGN	440	25	25	19	22	28	23	31	32
RADIO-AGN	1 141	24	25	19	24	27	24	31	32
SED-AGN Only	4 291	8	7	8	6	7	7	6	7

The ROC curve is sensitive to model assumptions (see App. § 8). For a subset with FLAGINFRARED = 4 and metallicity left as a free parameter, the performance is similar or even slightly better (AUC = 0.88) than for the default configuration (see also Tab. C.1).