The PHANGS-HST Survey: Physics at High Angular Resolution in Nearby Galaxies with the Hubble Space Telescope

For each of the 43 new fields, we aimed to cover the region mapped in CO(2–1) by ALMA in five filters: F275W (NUV), F336W (U), F438W (B), F555W (V), and F814W (I). F275W is the shortest wavelength filter that avoids the 2175 Å dust feature, and in combination with the U and B bands serves to break the age-extinction degeneracy. The V and I bands are less affected by extinction and variation in the mass-to-light ratio, and serve to constrain the stellar mass.

For 31 fields (in 26 galaxies), new observations with the WFC3 UVIS camera were needed in all five filters. Three exposures with subpixel dithering were taken in each filter (the average pixel size of WFC3 UVIS is 004), and all 15 exposures were obtained in a three-orbit visit, yielding total exposure times of ∼2200 s (NUV), ∼1100 s (U), ∼1100 s (B), ∼670 s (V), and ∼830 s (I). A post flash of 5–10 e⁻ is applied for each exposure, with larger values in the shorter wavelength filters, to increase the background to 12 e⁻, which is the recommended value to mitigate issues due to charge transfer efficiency (CTE) losses. The dither sequence is optimized to cover the WFC3 chip gap and to help recover the undersampled point-spread function (PSF).

For the other 12 pointings (in eight galaxies), suitable WFC3 or ACS data in one or more of these filters were taken by prior programs. The available archival data were obtained from MAST and processed in a consistent manner with our new observations, and new imaging was obtained only for the missing filters within a two-orbit visit.

While the primary observational goal of the PHANGS-HST survey is to obtain UV–optical imaging of the star-forming disk, we also simultaneously observe the galaxy halo with the Advanced Camera for Surveys Wide Field Channel (ACS/WFC) in "parallel" mode. Such parallel observations can potentially yield measurements of the galaxy distance if the tip of the red giant branch (TRGB) can be identified in a color–magnitude diagram of halo stars. Thus, we have designed our observations so that ACS imaging in F606W and F814W, filters commonly used for TRGB analysis (e.g., McQuinn et al. 2017; Anand et al. 2018) accompanies each WFC3 "prime" observation.

For the range of distances and angular sizes of the PHANGS-HST galaxies, the ACS field of view generally falls on the halo of the target galaxy when WFC3 is centered on the galaxy itself (Figure 3). Given that the science requirements of PHANGS-HST constrain the positioning of the prime pointings, optimizing placement of the parallel fields (as in a focused TRGB program) is a secondary priority and is restricted by the fixed spatial offset of the two cameras on the focal plane. For galaxies with relatively large angular sizes, the parallel observations may include portions of the outer disk. For smaller galaxies, the parallels may be too far to capture a significant number of halo stars. To the extent possible, ORIENT constraints were imposed to prevent the ACS field from entirely falling on the galaxy disk, on nearby galaxy neighbors, and/or on extremely bright foreground stars. For some targets with large angular sizes where it was not possible to avoid the disk completely, the field was positioned along the major axis to help differentiate between disk and halo stars. In several cases, the desired ORIENT constraints were lifted or relaxed to allow guide stars to fall into the area of the focal plane accessible to the Fine Guidance Sensors.

The five-band prime observations with WFC3/UVIS were sequenced in each orbit to optimize exposure time in the parallel observations without impacting the primary observations. As discussed in the previous section, WFC3 observations for each pointing required two or three orbits, depending on whether suitable HST archival observations of the galaxy disk were already available. For two-orbit visit pointings, the total exposure times in the ACS parallel V and I images are ∼2100 s each, while for the three-orbit visits they are about ∼3500 s and 3200 s, respectively. In both cases, three exposures were taken in each filter.

Distance constraints resulting from analysis of the TRGB based on parallel imaging obtained in the first year of the PHANGS-HST program (for 30 galaxies through 2020 July) are presented in Anand et al. (2020). We were able to measure TRGB distances for 10 PHANGS galaxies, 4 of which are the first published TRGB distance measurements for those galaxies (IC 5332, NGC 2835, NGC 4298, and NGC 4321) and 7 of which represent the best-available distances (IC 5332, NGC 2835, NGC 3621, NGC 4298, NGC 4826, NGC 5068, and NGC 6744). Analysis of the remaining six galaxies (with seven parallel fields: IC 1954, NGC 0685, NGC 1097, NGC 2903-N/S, NGC 5068, ⁴⁷ and NGC 7496) yields one additional TRGB constraint. A distance of 9.61 ± 0.39 Mpc is found for NGC 2903-N, and represents the best-available distance for that galaxy. All PHANGS-HST TRGB measurements and accompanying color–magnitude diagrams will be available from the Extragalactic Distance Database (EDD; Tully et al. 2009; Anand et al. 2021). ⁴⁸ In total, the PHANGS-HST ACS parallel observations have provided eight new TRGB measurements which are the current best-available distances. The eight measurements span from 4.41 ± 0.19 Mpc (NGC 4826) to 14.9 ± 1.4 Mpc (NGC 4298), and are listed in Table 1 together with other adopted distances from the literature compiled by Anand et al. (2020).

The process used to drizzle and mosaic the HST imaging data follows current standard procedures.

Data acquired for PHANGS-HST are first obtained from MAST, along with other suitable archival data taken by previous programs. These "FLT" exposures have been processed through the standard Pyraf/STSDAS CALACS or CALWFC3 software in the archive, which performs initial data quality flagging, bias subtraction, gain correction, bias stripe removal, correction for CTE losses, dark current subtraction, flat-fielding, and photometric calibration, resulting in "FLC" FITS files for each ACS/WFC and WFC3/UVIS exposure.

The PHANGS-HST pipeline is based on the STScI-supported software package DRIZZLEPAC to combine exposures and improve sampling of the PSF and is used for the prime observations targeting the galaxy disk in a two-step drizzle procedure. The parallel observations are treated separately as discussed in Anand et al. (2020).

The pipeline takes the FLC FITS files retrieved from MAST as input to produce combined images for each filter, which are all aligned and drizzled onto a common grid with a pixel scale of 004 (the native WFC3 pixel scale), with astrometry calibrated with GAIA DR2 sources (Gaia Collaboration et al. 2018; Lindegren et al. 2018). The latter is essential for the proper alignment of the HST imaging with ALMA and VLT/MUSE data and to facilitate a joint study of the three data sets.

The V-band imaging (WFC3 F555W) is used as the reference for the positioning of the images in all other filters (NUV, U, B, and I) and to define the common pixel grid. Using the F555W FLC files from MAST for each pointing, the sky positions of the centers and corners of the images are calculated to define a search area to query the ESA DR2 GAIA catalog. The TWEAKREG routine then matches the GAIA sources to the objects detected in the F555W drizzled image and calculates average shifts (with an accuracy typically better than 0.1 pixel) to correct the astrometric solution. The number of GAIA sources found in a given F555W HST pointing varies from as few as 15 sources to a maximum of 317 with an average of 40 sources. The TWEAKBACK routine is then used to propagate the corrected WCS solution back to the original F555W FLC images. Finally, TWEAKREG and TWEAKBACK are used again, but now with the drizzled images for the other four filters to find sources in common with those detected in F555W and to align to the F555W image.

The final drizzle combination is carried out with sky subtraction using the "globalmin+match" method for sky calculation. ASTRODRIZZLE first finds a minimum "global" sky value for each chip/image extension in all input images and then uses the "match" method to compute differences in sky values between images in common sky regions and to equalize the sky values between images. The final DRC FITS files are in units of e⁻ s⁻¹ and are registered with north up and east left as usual. "EXP" (exposure time) and "ERR" (error) weight (WHT) maps, calculated from the contribution of all input exposures to a given output pixel, are also produced. The ERR maps includes all noise sources from detector and sky and are used to compute uncertainties in photometry downstream in the pipeline.

Color composites of the drizzled, mosaicked HST imaging are shown in Figure 4 for the same six galaxies featured in Figures 2 and 3.

Zoom In Zoom Out Reset image size

The principal method of source detection in the PHANGS-HST pipeline is provided by the DOLPHOT photometry package (v2.0; Dolphin 2016), which is based on PSF fitting and operates on the FLC files to detect and deblend sources in HST imaging. We elect to use DOLPHOT for source detection in order to provide a common starting point for the identification of both single-peaked compact star clusters, and multipeaked stellar associations over a range of physical scales. The drizzled V-band image is used as the positional reference and sources are detected to 3.5σ with PSF fitting performed at the same image positions in all bands.

Compact star clusters have effective radii between 0.5 pc to about 10 pc (Portegies Zwart et al. 2010; Ryon et al. 2017). At the distances of the galaxies in the PHANGS-HST sample, such objects will have angular sizes close to the HST WFC3 resolution (i.e., 2 pixels, ∼008; from 1.7 to 9 pc for the range of distances for the galaxy sample) and will appear sufficiently point-like to be captured by DOLPHOT. To ensure that the catalogs include star clusters with light profiles that may be more extended than the sources detected by DOLPHOT, source detection is also performed on the V-band DRC image with DAOStarFinder (the Python implementation of DAOPHOT within the photutils astropy-affiliated package), with a kernel FWHM of 2.5 pixels and an effective signal-to-noise ratio (S/N) consistent with the detection threshold used for DOLPHOT. The merged DOLPHOT and DAOStarFinder catalogs provide the source lists from which compact star cluster candidates are identified, while the branch of the pipeline that identifies multipeaked stellar associations (Section 4.7) relies only upon the DOLPHOT catalogs, as indicated in the flowchart in Figure 5.

Zoom In Zoom Out Reset image size

Following source detection, the star cluster pipeline then uses the positions from the merged DOLPHOT-DAOStarFinder catalog to perform aperture photometry in all five filters. For all galaxies photometry is measured in circular apertures with a 4 pixel radius, with the background determined in an annulus between 7 and 8 pixels around the aperture. An aperture correction (Section 4.3) is applied that yields the total fluxes for sources ultimately identified as compact star clusters. Photometry is also measured in a series of circular apertures (with radii from 1 to 5 pixels) to compute concentration indices that are used to distinguish star cluster candidates from stars and other sources/artifacts (Section 4.4). Catalogs with these observed parameters are produced for each galaxy, which on average contain about ∼500,000 sources. The contribution of nonredundant detections from DAOStarFinder to these catalogs is less than ∼1% of objects with a V-band aperture photometry S/N > 10, the limit subsequently applied to select star cluster candidates as discussed in Section 4.4.

Thilker et al. (2022) discuss in detail the parameter choices for both the DOLPHOT and DAOStarFinder routines, and the procedures for obtaining aperture photometry and computing errors.

Star clusters, in particular those that are young, are generally found in crowded regions. Direct, accurate measurement of the total flux is often not possible since the outer light profile is frequently contaminated by other sources. Thus, HST photometry of compact star clusters in galaxies beyond the Local Group is typically measured with a limited aperture that captures ∼50% of the total flux, and then a correction, determined from bright, isolated clusters, is applied (e.g., Chandar et al. 2010b; Adamo et al. 2017; Cook et al. 2019).

Deger et al. (2021) present a detailed discussion of procedures used to determine average aperture corrections and uncertainties for each galaxy. In summary, the V-band images are visually inspected to identify a few dozen well-detected, isolated, compact clusters, and these objects are used to compute an average correction for each field. Fixed offsets, based on the change in the encircled energy distributions of the WFC3 PSF with wavelength (which has a minimum FWHM of 0067 in the V band, and increases to ∼0075 in the NUV and I bands) ^{49 , 50} are used to calculate the corresponding corrections for photometry in the NUV, U, B, and I bands, since direct measurements from growth curves in those filters for bright sources in the V band that are very red or blue can be noisy (Cook et al. 2019; Deger et al. 2021). By construction, the resulting V-band corrections are ∼0.75 mag (for the adopted photometric aperture radius of 4 pixels, i.e., 016). The corrections are larger by 0.19, 0.12, 0.03, and 0.12 mag for the NUV, U, B, and I bands, respectively.

The PHANGS-HST pipeline identifies cluster candidates from the source lists described in Section 4.2 using photometric and morphological properties measured in the V band. In previous work, the concentration index (computed as the difference between photometry measured in circular apertures with radii of 1 and 3 pixels, CI₁₃ or simply CI) has been generally used to remove sources likely to be stars from consideration (e.g., Whitmore et al. 2010; Adamo et al. 2017; Cook et al. 2019). To determine the threshold to separate stars from cluster candidates, CIs have been typically measured for a few dozen isolated, bright point sources in an image. These measurements are then compared with the CI distribution for an analogous sample of compact star clusters (identified through visual inspection and also used to derive aperture corrections).

For PHANGS-HST, Thilker et al. (2022) build upon this method to develop candidate selection criteria based on multiple concentration indices (MCI), rather than a single concentration index. We measure fluxes in circular apertures with radii of 1.0, 1.5, 2.0, 2.5, 3.0, 4.0, and 5.0 pixels, and we define two metrics to characterize the light profiles between 1 and 2.5 pixels (MCI_in) and 2.5–5 pixels (MCI_out), where

$$\begin{eqnarray}&&\mathrm{MCI}\equiv \displaystyle \frac{1}{3}({\mathrm{NCI}}_{{ab}}+{\mathrm{NCI}}_{{bc}}+{\mathrm{NCI}}_{{cd}}),\end{eqnarray} \tag{ 1 }$$

where a, b, c, and d represent the radii of the apertures, and

$$\begin{eqnarray}&&{\mathrm{NCI}}_{{ij}}\equiv 1-\displaystyle \frac{{\mathrm{CI}}_{{ij}}}{{\mathrm{CI}}_{{ij},\mathrm{fiducial}}}\end{eqnarray} \tag{ 2 }$$

is a CI normalized by CI_ij,fiducial, defined to be the CI of a relatively compact cluster; i.e., a PSF-convolved Moffat/EFF87 function (Moffat 1969; Elson et al. 1987, EFF87) with a FWHM of 2 pixels, and a power-law slope of 3 describing the surface brightness profile of the extended halo. The choice of normalization is arbitrary, and enables the various CIs to be meaningfully averaged. With these definitions, clusters measured on HST optical images generally have values of −0.5 ≲ MCI_in ≲ 0.2, −2 ≲ MCI_out ≲ 0.6 (see Figures 7 and 6). MCI_in is anticorrelated with the standard CI₁₃, as would be expected from their definitions; i.e., more compact sources are characterized by larger MCI values (i.e., more concentrated), which is the opposite of the sense of the standard CI₁₃. By construction, the fiducial cluster lies at the origin of the MCI_in–MCI_out plane.

We define selection regions in two ways on the MCI_in–MCI_out plane (Figure 6) to generate two types of candidate lists. (1) The first set of selection regions are contours in the MCI plane, based on the loci of synthetic star clusters inserted into the V-band imaging for each individual target galaxy. This selection strategy yields up to several thousand candidates per target, and CNN models are used for their classification. (2) The second are smaller polygonal regions, the same set of which are used for all galaxies. Definition of these polygons is based on the loci of human inspected and verified clusters, together with the inner synthetic star cluster contours (which encloses the highest concentration of models) for the first few PHANGS-HST galaxies studied. This second selection produces smaller samples, more suitable for time-intensive, expert human classification; it is intended to yield an average of ∼1000 candidates per field, so that inspection over the full set of 38 PHANGS-HST galaxies is manageable. Once defined, it provides a simpler, faster way of selecting cluster candidates for inspection by bypassing the synthetic star cluster analysis required to generate contours for a given galaxy. More specifically:

• 1.  
  Selection criteria in the MCI plane are defined based on model star clusters. Synthetic clusters are generated with Moffat profiles, which again are parameterized by the effective radius (in units of the intrinsic/pre-PSF-convolution FWHM), and the power-law slope describing the surface brightness profile of the extended halo. The modeling includes 216 distinct Moffat profiles, spanning 0.5 pix ≤FWHM ≤ 7 pix and power-law slopes from 0.75 to 4. Models are generated for each of these profile types with a distribution of ∼4000 apparent magnitudes. The magnitudes are computed from the distance of the galaxy and V-band luminosities based on solar metallicity, single-aged stellar population models of Bruzual & Charlot (2003), for a grid of masses (10³–10⁵ M_⊙), ages (1–1000 Myr), and extinctions (0 ≤ A_V ≤ 0.5 mag). The synthetic clusters are randomly inserted into the V-band images 200 at a time, and aperture photometry is performed with the same procedure used to measure real sources. In total, ∼4 × 10⁶ clusters are inserted. The MCI values from these synthetic clusters are plotted on the MCI plane to form a 2D histogram with a bin size of 0.01, and a contour is derived. Contours are also generated for MCI bin sizes of 0.02 and 0.04, resulting in three nested model selection regions for each field (Figure 6). Typically, the inner model region contains the majority of candidates (∼50%–70%), while the two larger contours each add comparable numbers of candidates (20%–30%). Currently, the outer contour (i.e., corresponding to the largest bin size) is used to produce candidate lists for automated classification by CNNs (see Section 4.5).
• 2.  
  Selection criteria in the MCI plane are also defined semi-empirically. For the first few PHANGS-HST galaxies studied (e.g., NGC 1559, Wei et al. 2020), sources were inspected over a wide swath of the MCI plane (essentially spanning the outer model contour). The loci of human verified clusters were used in combination with the tightest model contour to define the left, right, and top edges of the polygon. Catalogs of visually inspected clusters published by the HST LEGUS program for 34 HST fields in 30 galaxies ⁵¹ were then used to verify that the majority of LEGUS clusters would be captured by this selection (Figure 7). The lower bound of MCI_out increases as the distance of the host galaxies increases—clusters will be less resolved and appear more compact at larger distances, and both the model contours and the loci of the human verified cluster populations do indeed shrink with increasing distance. Using the ensemble of cluster populations identified in the first few PHANGS-HST galaxies studied, plus those in the LEGUS catalogs, we adopt limits of MCI_out = −1.7, −1.1, and −0.7 for galaxies with distances ≤8 Mpc, 8 Mpc <d < 14 Mpc, and ≥14 Mpc, respectively (Figure 7). This initial analysis over a wide swath of the MCI plane for PHANGS-HST sources suggests that the density of clusters rapidly drops beyond the boundaries of the polygon, but a more careful investigation of such completeness issues will be the subject of future work. These semi-empirical polygon selection regions are used to produce candidate lists for human inspection and classification (see Section 4.5 and Whitmore et al. 2021) to a total V-band limit of ∼24 mag (corresponding to absolute V-band limits between about −5.5 mag and −8 mag for the distances of the galaxies in the sample). The exact value of the magnitude limit depends on the number of candidates, since, again, the primary purpose of this second selection is to reduce the number of candidates for visual inspection to a manageable level (i.e., ≲1500 per galaxy) while maximizing the yield of true clusters. So, if the total number of candidates is low (several hundred) then the limit is fainter, whereas if the number of candidates is large, the limit is somewhat brighter.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

For both model-based (contours) and semi-empirical (polygon) selection methods:

• 1.  
  The regions in the MCI plane dominated by stars (point sources) as detected by DOLPHOT, are defined for each field and used to exclude objects. Figure 6 shows both the synthetic and semi-empirical cluster selection regions for four galaxies at a range of distances together with their stellar exclusion regions.
• 2.  
  Candidates must also satisfy basic criteria: V-band photometry measured in a 4 pixel radius must have S/N ≥ 10, and the source must also be detected in at least two other bands with photometric error ≤0.3 mag. The faintest sources in the resulting candidate lists for fields with the standard V-band exposure time of 670 s have total V-band magnitudes of ∼24.6, which corresponds to absolute magnitudes between −4.2 and −7.8 for the distance range spanned by the PHANGS-HST galaxies.
• 3.  
  All sources with total, absolute V-band magnitudes brighter than −10 mag and −0.55 ≤ MCI_in ≤ 0.45 (i.e., MCI_in values plausible for real objects) are kept as candidates to help ensure high completeness for the brightest clusters. Sources brighter than −10 mag exceed the Humphreys–Davidson (HD) limit, the observed maximum luminosity of individual stars in the LMC, thought to be due to a modified Eddington Limit (Humphreys & Davidson 1979; Lamers & Levesque 2017). These "super HD sources" also undergo classification to remove interlopers such as saturated stars and background galaxies.

The number of candidates identified using the empirical MCI selection regions varies from many hundred to several thousand sources for each field, with a median of ∼1000 (Figure 8), and should ultimately yield sample sizes of compact star clusters similar to those studied in previous work. The variation in candidate sample sizes reflects the variation in global sSFRs for the PHANGS-HST galaxies that span a factor of ∼10 (Figure 1). The synthetic cluster MCI selection casts a wider net for potential clusters (Figure 6) and yields candidate samples about a factor two larger than the empirical selection (Figure 8). These larger samples enable analysis of potential incompleteness in previous star cluster studies, in particular for more diffuse clusters that appear to be rare.

Zoom In Zoom Out Reset image size

These new selection methods, based on the measurement of multiple CI and the use of model star clusters, provides a solid foundation for quantitative investigation of structural properties including: which model clusters actually exist in nature, whether certain clusters are likely to be bound or unbound, and how their morphologies evolve with time. Our model grid of clusters also facilitates future work to characterize cluster completeness levels (e.g., Johnson et al. 2015). More discussion on the utility of the approach to cluster selection is provided in Thilker et al. (2022).

The candidate star clusters then undergo a process of inspection to further remove contaminants and sort the clusters into different morphological categories. A combination of human visual inspection and automated inspection with CNN models is performed as follows:

• 1.  
  The smaller lists of candidates resulting from the semi-empirical MCI selection region (polygons) are visually inspected by co-author B.C.W. Objects with a total V-band magnitude to a limit of ∼24 mag receive visual classifications.
• 2.  
  The samples of candidates identified in the largest synthetic cluster MCI selection region (contours) are classified by CNN models, as described below. Candidates identified using the semi-empirical selection region generally lie within this synthetic cluster selection area (Figure 6), and hence will have both neural network and human visual classifications.
• 3.  
  Human visual inspection of fainter candidates and those beyond the boundaries of the empirical selection region are performed on an ad hoc basis to evaluate and monitor the performance of the neural network models.

For PHANGS-HST, we adopt the general classification scheme used by LEGUS as described in Adamo et al. (2017) and Cook et al. (2019):

• 1.  
  Class 1: compact star cluster—single peak, circularly symmetric, but radial profile more extended relative to point source.
• 2.  
  Class 2: compact star cluster—similar to Class 1, but elongated or asymmetric.
• 3.  
  Class 3: compact stellar association—asymmetric, multiple peaks.
• 4.  
  Class 4: not a compact star cluster or compact stellar association (e.g., image artifacts, background galaxies, individual stars or pairs of stars).

Examples of Class 1, 2, and 3 objects are shown in Figure 9. Whitmore et al. (2021) provide a detailed description of the process of visual inspection and morphological classification, and discuss differences in the application of this scheme for Class 3 objects relative to the LEGUS project (i.e., we require evidence of four or more peaks within a radius of 5 pixels in PHANGS-HST). A brief history of star cluster classification is also given there (also see Wei et al. 2020, Section 2). While we continue to include Class 3 objects in our compact cluster catalogs, we note that this is mainly for historical continuity with the LEGUS project. The PHANGS-HST pipeline is optimized to identify single-peaked compact clusters, and this leads to a high level of incompleteness for multipeaked stellar associations. Instead, we introduce a new identification process for stellar associations, based on a watershed algorithm, which provides a far more complete inventory of young stellar populations and the star formation hierarchy at multiple physical scales, as will be summarized in Section 4.7 and presented in K. Larson et al. (2021, in preparation).

Zoom In Zoom Out Reset image size

We note that it is debated whether such classifications for a given object can distinguish between gravitationally bound clusters and unbound associations, which may form and evolve under distinct conditions (Gieles & Portegies Zwart 2011; Kruijssen 2012; Ryon et al. 2017; Krumholz et al. 2019; Ward et al. 2020). Statistically however, it is expected that Class 1 should contain the highest percentage of bound clusters, and Class 3 should have the highest percentage of unbound associations (see Whitmore et al. 2021, for further discussion).

Ultimately, PHANGS-HST will generate up to ∼80,000 star cluster candidates for inspection and classification. In previous large studies of star clusters, the process of visual inspection has been a limiting step, which motivated the investigation of automated machine-learning techniques (Messa et al. 2018; Grasha et al. 2019; Perez et al. 2021). In Wei et al. (2020), we studied the application of deep transfer learning techniques to train CNN to classify star cluster candidates according to the scheme above. Deep transfer learning involves the tuning of a pretrained network, for example, based on the ImageNet library of everyday objects. ⁵² In principle, this approach enables CNNs to be successfully trained with relatively small samples (i.e., hundreds to a thousand images), which is the current size of HST star cluster samples with visual classifications. It provides an alternative to the process of training all network layers from scratch, which requires samples that are an order of magnitude larger. The results of Wei et al. (2020) were encouraging, as the prediction accuracies (70%, 40%, 40%–50%, and 50%–70% for Class 1, 2, and 3 star clusters and Class 4 nonclusters, respectively) were found to be competitive with the classification consistency between different human classifiers. The neural network models presented in Wei et al. (2020) provide a starting point for automated classification of the PHANGS-HST and other HST star cluster candidate samples, which can continue to be optimized. Whitmore et al. (2021) present results of the current Wei et al. (2020) models applied to clusters candidates in five PHANGS-HST galaxies, which also have classifications published by the LEGUS project. The Whitmore et al. (2021) analysis includes a detailed comparison between human and automated classifications, from overall prediction accuracy to differences in the distribution of ages, UBVI color–color diagrams, and stellar mass functions, and also includes a discussion of additional future work to improve performance.

To derive ages, masses, and reddenings for the sources classified as star clusters and associations, we use a modified version of cigale ⁵³ (Code Investigating GALaxy Emission; Burgarella et al. 2005; Noll et al. 2009; Boquien et al. 2019), a publicly available SED-fitting package developed for galaxies. Our modifications, which support the fitting of single-age populations and provide modeling options to facilitate comparison to prior SED modeling results for star clusters, are available in dedicated branches, SSP and SSPmag respectively, of the public git repository ⁵⁴ of cigale. Turner et al. (2021) reports on these modifications and the analyses performed to validate the code.

SED fitting with cigale is performed on the five-band photometry (NUV–U–B–V–I) for both compact star clusters and stellar associations. The fitting is based on the simple (single-aged) population synthesis models of Bruzual & Charlot (2003), assuming solar metallicity and a Chabrier (2003) IMF (with standard mass limits of 0.1–100 M_⊙), and no addition of nebular emission. The Cardelli et al. (1989) extinction curve with R_V = 3.1 is used and a maximum E(B − V) = 1.5 mag is imposed. The reasoning for these choices is discussed in Turner et al. (2021).

When CIGALE is run with the same assumptions and theoretical models as used in LEGUS, Turner et al. (2021) find very similar cluster ages and masses. Quantitatively, fits to identical cluster photometry yield (logarithmic) medians in the ratios of ages and masses for the two surveys of 0.001 ± 0.017 dex and 0.003 ± 0.011 dex, respectively.

The majority of star formation occurs in stellar associations (Lada & Lada 2003; Ward & Kruijssen 2018; Ward et al. 2020; Wright 2020, and references therein). Compact star clusters, the focus of the previous sections, are formed only in the densest peaks of the star formation hierarchy (Elmegreen 2008; Kruijssen 2012) and contain between 1% ∼ 50% of the total star formation in galaxies (Kruijssen 2012; Adamo et al. 2015, 2020; Johnson et al. 2016; Chandar et al. 2017; Krumholz et al. 2019). To produce catalogs of stellar associations, methods distinct from those used to identify single-peaked compact clusters are needed to segment the light distribution over larger physical scales and to probe further into the star formation hierarchy. Development of such methods is particularly important for obtaining complete inventories of the youngest stellar populations (≲10 Myr), which are a requirement for a robust joint analysis with molecular clouds and H ii regions.

For PHANGS-HST, K. Larson et al. (2021, in preparation) develop a technique to produce catalogs of stellar associations spanning scales from 8 to 64 pc. The technique builds upon the watershed routine in the scikit-image Python package (skimage.segmentation.watershed, van der Walt et al. 2014), which is based on the concept of geological watersheds. The routine identifies regions by "flooding" an image, given a set of markers as the starting points. In addition to the image on which associations are to be identified (i.e., requiring segmentation), the inputs needed by watershed are a list of marker positions and an image mask that defines the areas over which regions are allowed to grow.

Our adopted technique deploys watershed on a smoothed, filtered map of the positions of point sources (rather than directly on the HST images) and uses a two-parameter procedure to determine the marker positions and to produce the image mask from these position maps. The smoothed, filtered, position maps are produced as follows. Point sources are selected from the DOLPHOT catalogs that satisfy basic requirements on signal-to-noise ratio, sharpness, and data quality. The positions of these point sources are used to create maps with the same pixel grid as the PHANGS-HST DRC images, containing values of 1's corresponding to the DOLPHOT positions, and 0's otherwise. The maps are then smoothed with Gaussian profiles with FWHM of 2ⁿ pc for n = 3, 4, 5, and 6 (i.e., 8, 16, 32, and 64 pc), computed given the distance of the galaxy. Finally, a high-pass filter is applied by subtracting a map that has been smoothed with a kernel that is four times larger.

Two parameters are then defined that are tied to the characteristics of a single object on these smoothed, filtered position maps. The peak threshold parameter is the level above which markers (local maxima) are identified; it is currently defined to be 1.5 times the maximum value for a single object so that the resulting regions are multipeaked. The edge threshold parameter is the minimum surface "brightness" level beyond which the regions are not allowed to expand further, and used to create a mask image for watershed; it is defined to be the surface "brightness" at the FWHM of a single object.

The smoothed, filtered, position maps enable the identification of structures on the physical scales over which the maps have been smoothed. The 8 pc smoothed maps allow associations that overlap in size with sources in the compact cluster catalog to be studied, and the maps that are smoothed over larger scales enable the greater star formation hierarchy to be traced. Structures are identified based on the NUV and V-band DOLPHOT point-source catalogs. The resulting NUV-band selected association catalogs will predominantly contain young structures (≲100 Myr). In comparison, V-band selected catalogs will include structures over a larger range of ages and will facilitate comparison with the compact cluster samples, which have also been V-band selected.

K. Larson et al. (2021, in preparation) demonstrate the validity of this technique for identification of stellar associations in the PHANGS-HST galaxy sample based on an analysis of NGC 3351 and NGC 1566.

Figure 9 shows the watershed-identified stellar associations in NGC 3351 together with visually classified objects from the compact star cluster pipeline (Section 4.4). Three different magnifications of the galaxy are shown to illustrate the full disk and star-forming ring (bottom panels), the multiscale associations traced by the watershed method (top left and middle panels), and individual compact star clusters and associations (series of postage stamps at top right). The structures are overlaid on color composites of the HST imaging, ALMA CO(2–1) map, and a color composite of the Hα and [N ii] λ6583 maps from MUSE. Only the youngest (<10 Myr) watershed associations are shown to illustrate the correlation with the blue starlight, molecular clouds, and H ii regions. Ages are derived using CIGALE as summarized in Section 4.6 and documented in Turner et al. (2021). Details on the procedures used to measure fluxes in the regions are presented in K. Larson et al. (2021, in preparation).

From examination of the properties of the resultant watershed structures, we find:

• 1.  
  Sample sizes of several hundred up to a few thousand associations in each galaxy. For associations identified on physical scales comparable to the aperture sizes used for selecting compact clusters candidates (a 4 pixel radius, which corresponds to 8 pc at a distance of 10 Mpc; Section 4.4), the numbers of associations identified are about ∼2–4 times larger than the numbers of visually classified compact clusters.
• 2.  
  The process is shown to successfully identify structures at the defined scale. The size distributions are well defined, and approximately log-normal with medians near the FWHM of the smoothing kernel.
• 3.  
  Fluxes computed within the boundaries of regions identified on the 8 and 16 pc smoothed images yield colors that are consistent with single-aged stellar population tracks on the UBVI color–color diagram. The loci are similar to previously studied samples of compact star cluster and associations (Figure 10). The similarity between two panels of Figure 10 demonstrates that these structures, which have been identified with a somewhat complicated algorithm, do behave as one might expect for groups of stars that are physically associated and born close in time. The associations may not strictly be singled-aged populations, but rather composite populations that can be approximated by a single-aged model. The level of deviation from a single-aged population, particularly as a function of the size of the region, is an issue that will be the topic of future work. Such work will be facilitated by the PHANGS-HST pipelines use of CIGALE, which allows for the self-consistent modeling of both single-aged and composite stellar populations as discussed in Turner et al. (2021).
• 4.  
  Maps of the youngest associations (<5 Myr), show excellent correspondence to H ii regions observed in narrowband Hα imaging. As the age of the regions increases, they become more anticorrelated with the H ii regions, as would be expected (Figure 11). This provides some evidence that the age-dating of the youngest structures appears to produce reasonable results, despite the simplifying assumption of a single-aged population in the SED modeling and the complexities of the overall process to define the regions and measure their photometry.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Initial testing of our watershed-based methods for identifying stellar association have been performed with NGC 3351 and NGC 1566. These two galaxies were chosen as they span a significant range of distances in the sample (from ∼10–18 Mpc), and because catalogs of clusters and compact associations are available from LEGUS for comparison. As with our new methods for identifying clusters, the completeness of our samples of multiscale associations, particularly possible systematics as a function of distance, will need to be studied with simulations.

For imaging of the star-forming disk in NUV–U–B–V–I bands for the 38 PHANGS-HST galaxies:

• 1.  
  FLC FITS file for each exposure with astrometric solutions updated based on GAIA DR2 sources.
• 2.  
  Combined DRC FITS images of individual pointings in each filter, each drizzled onto a common pixel grid defined for the galaxy target, also with astrometric solutions based on calibration to GAIA DR2 sources.
• 3.  
  Mosaicked DRC FITS images in each filter for 13 galaxies covered by multiple pointings (NGC 628, NGC 1097, NGC 1300, NGC 1512, NGC 1672, NGC 2903, NGC 3351, NGC 3621, NGC 3627, NGC 4254, NGC 4321, NGC 4536, an NGC 6744).
• 4.  
  ERR and EXP weight FITS images for individual pointings as well as mosaics.

• 1.  
  DOLPHOT catalogs with five-band PSF-fitting photometry.
• 2.  
  Compact star cluster and stellar association candidate catalogs, including position, five-band aperture photometry, stellar mass, age, reddening, CNN morphological classification, visual classifications for a subset of candidates in the empirical selection region, multiple concentration indices (MCI), and standard concentration index (CI) values.
• 3.  
  Catalogs of stellar associations detected at 8, 16, 32, and 64 pc scales, including five-band region photometry, stellar mass, age, reddening, and effective radius, together with DS9 region files providing peak position and boundaries of regions, and FITS masks of regions.

• 1.  
  The Python routines that constitute the PHANGS-HST compact star cluster and association pipeline will be released at https://github.com/PhangsTeam.
• 2.  
  CIGALE augmentations for SED fitting of single-aged stellar populations are available in dedicated branches, SSP and SSPmag, respectively, of the public git repository ⁵⁶ of cigale.
• 3.  
  CNN models for cluster candidate classification as described in Wei et al. (2020) and Whitmore et al. (2021). An annotated Python notebook containing scripts to run the models will be provided. Future updates shown to be improvements over the current models will be released as they are developed.

PHANGS-ALMA data have been released for the full PHANGS parent sample through the ALMA Archive and the Canadian Astronomy Data Centre (CADC ⁵⁷ ), and also linked to the PHANGS portal at MAST. The PHANGS-ALMA products include the ¹²CO(2–1) spectral line data cubes, signal masks, and derived products such as the integrated intensity, line-of-sight velocity estimate, and spectral line widths. A description of the data reduction, products, and release is provided in Leroy et al. (2021a).

Likewise, PHANGS-MUSE integral field spectrograph data of the 19 galaxies targeted in the course of VLT Large Programme (ESO 1100.B-0651) has been released via the ESO Science Archive ⁵⁸ and CADC, and also linked to the PHANGS portal at MAST. The released MUSE data include reduced and fully mosaicked data cubes as well as a series of two-dimensional maps associated with the gas and stellar tracers: broadband reconstructed images, emission line distribution and kinematics, stellar kinematics, star formation histories (mass and light-weighted age and metallicity maps), extinction maps from Balmer decrement, and stellar continuum fitting. The details of the data reduction and analysis processes are provided in Emsellem et al. (2021).

Links to the archive locations for all released PHANGS products are available at the survey webpage (http://phangs.org/data).