Photometrically-Classified Superluminous Supernovae from the Pan-STARRS1 Medium Deep Survey: A Case Study for Science with Machine Learning-Based Classification

With the upcoming Vera C.~Rubin Observatory Legacy Survey of Space and Time (LSST), it is expected that only $\sim 0.1\%$ of all transients will be classified spectroscopically. To conduct studies of rare transients, such as Type I superluminous supernovae (SLSNe), we must instead rely on photometric classification. In this vein, here we carry out a pilot study of SLSNe from the Pan-STARRS1 Medium-Deep Survey (PS1-MDS) classified photometrically with our SuperRAENN and Superphot algorithms. We first construct a sub-sample of the photometric sample using a list of simple selection metrics designed to minimize contamination and ensure sufficient data quality for modeling. We then fit the multi-band light curves with a magnetar spin-down model using the Modular Open-Source Fitter for Transients (MOSFiT). Comparing the magnetar engine and ejecta parameter distributions of the photometric sample to those of the PS1-MDS spectroscopic sample and a larger literature spectroscopic sample, we find that these samples are overall consistent, but that the photometric sample extends to slower spins and lower ejecta masses, which correspond to lower luminosity events, as expected for photometric selection. While our PS1-MDS photometric sample is still smaller than the overall SLSN spectroscopic sample, our methodology paves the way to an orders-of-magnitude increase in the SLSN sample in the LSST era through photometric selection and study.

Several mechanisms have been proposed to power SLSNe, but a magnetar central engine model (Kasen & Bildsten 2010;Woosley 2010;Dessart et al. 2012;Metzger et al. 2015;Nicholl et al. 2017b) has had the most success in explaining both the light curves and spectra of the SLSN population. This model accounts for the broad range of peak luminosities and timescales (e.g., Nicholl et al. 2017b;Blanchard et al. 2020), for the early UV/optical spectra (e.g., Nicholl et al. 2017a), for the nebular phase spectra (e.g., Nicholl et al. 2016bNicholl et al. , 2019Jerkstrand et al. 2017), and for the power law decline rates observed in SN 2015bn and SN 2016inl at 10 3 d Blanchard et al. 2021). Additional support for a magnetar engine comes from the low metallicity host galaxies of SLSNe, which most closely resemble the hosts of long-duration gamma-ray bursts, another rare population of CCSNe that are likely powered by a central engine (Lunnan et al. 2014;Perley et al. 2016). While the magnetar engine model can explain the plethora of SLSN properties, other mechanisms have also been proposed to explain some SLSN properties; for example, Chen et al. (2022) recently argued that the light curves of at least some SLSNe from the Zwicky Transient Facility (ZTF; Bellm et al. 2019) can be explained equally well with a combination of circumstellar interaction (CSM) and Ni 56 decay. Furthermore, Hosseinzadeh et al. (2021) also explored ejecta-CSM interaction as a potential source for post-peak undulations in SLSN light curves.
With ongoing and upcoming wide-field optical surveys, including in particular the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST; Ivezić et al. 2019), only a small fraction of SNe are being classified spectroscopically (∼ 10% currently, and ∼ 0.1% anticipated for LSST; Villar et al. 2020). This impacts the ability to advance the study of rare SN classes, such as SLSNe, in particular. As shown by Villar et al. (2018), LSST may yield ∼ 10 4 SLSNe per year to z ∼ 3 (of which at least ∼ 20% will have well measured physical properties), but identifying these events requires photometric classification.
Recently, we presented two machine learning-based SN photometric classification pipelines, SuperRAENN  and Superphot , trained on 2315 SN-like transients from the Pan-STARRS1 Medium Deep Survey (PS1-MDS; Huber et al. 2017). Both classifiers use multiple SN classes, including in particular SLSNe. SuperRAENN combines a novel unsupervised recurrent autoencoder neural network (RAENN) with a random forest classifier for a semi-supervised algorithm. Superphot utilizes a random forest approach based on flexible analytic model fits to the light curves and their resulting parameters.
Here, as a demonstration of the type of approach and analysis that will be essential in the LSST era, we explore and study for the first time, the photometricallyclassified SLSNe from the Pan-STARRS1 Medium Deep Survey (PS1-MDS, Huber et al. 2017), as identified by SuperRAENN and Superphot. We first explore how to effectively construct a pure and well-measured subset of SLSNe from a photometrically-classified sample ( §2). We then model the light curves of the photometricallyclassified SLSNe with the same magnetar engine model previously used to study spectroscopically-classified SLSNe (using MOSFiT, Guillochon et al. 2018;§3). Finally, we compare the resulting parameter distributions to those of the spectroscopically-classified PS1-MDS SLSNe, as well as to the overall sample of spectroscopically-classified SLSNe ( §4).
Throughout the paper, we assume a flat ΛCDM cosmology with Ω m = 0.308 and H 0 = 67.8 km s −1 Mpc, based on the Planck 2015 results (Planck Collaboration 2016). We correct all photometry for Milky Way extinction using Schlafly & Finkbeiner (2011) and follow the extinction law of Fitzpatrick (1999) with R V = 3.1.

SAMPLE CONSTRUCTION
The data used in this paper are from the PS1-MDS. We refer the reader to Chambers et al. (2016) for details of the PS1 survey telescope and PS1-MDS observing strategy, and to Villar et al. (2020) and Hosseinzadeh et al. (2020) for the definition of the overall sample of SN-like transients and their light curves, description of the sub-sample of spectroscopically-classified events, the photometric classification approaches and results, all relevant data (including photometry and host galaxy redshifts), and complete descriptions of the algorithms and training processes.
In this paper we focus on the sample of photometrically-classified SLSNe. 1 Using SuperRAENN  and Superphot Hosseinzadeh et al. 2020) we photometrically classified 58 and 37 SLSNe, respectively, using the same training set of 557 spectroscopically-classified SNe, which includes 17 SLSNe that were studied in Lunnan et al. (2018). Here, we adopt the class with the highest probability as the predicted SN type for each transient.
Combining all transients classified by the two algorithms as SLSNe, and accounting for 28 classified as SLSNe by both, we obtain an initial sample with 67 photometrically-classified SLSNe. To further evaluate and potentially cull the photometric sample, we investigate several post-classification selection criteria. We find three effective criteria that help to reduce the sample contamination and lead to events with sufficient data to enable robust modeling. Furthermore, we apply an additional quality cut post-modeling based on model convergence. The criteria and their effects on the sample size are summarized in Table 1, and we discuss them in detail below.  Prior to applying our algorithms to the sample of PS1-MDS SN-like transients, we systematically excluded light curves with long-term variability to avoid contamination from active galactic nuclei (AGN). Still, some large AGN flares with little other variability over the 4.5 year time-span of the survey could survive this preliminary qualitative cut and eventually be classified as SLSNe. In particular, Hosseinzadeh et al. (2020) find that 14 photometrically-classified SLSNe with host galaxy spectra that exhibit broad AGN lines are located within 1 of the host center 2 . While these could in principle be SLSNe located indistinguishably close to an AGN, they are more likely large AGN flares or tidal disruption events, neither of which is a classification category in SuperRAENN and Superphot. Eliminating these events results in a combined sample of 53 events (Table 1, row 2).

Classification Confidence
Our initial sample requires that the highest classification probability be assigned as SLSN. However, given the number of classification categories, this does not necessarily mean that the classification confidence is high. Hosseinzadeh et al. (2020) and Villar et al. (2020) show that increasing the classification confidence threshold to p 0.75 leads to higher purity 3 across the full range of classes, at the expense of sample completeness. Here we apply a classification confidence threshold of p SLSN ≥ 0.5 as a compromise between purity and sample size (which corresponds to a purity of ≈ 0.78, see 2 These transients are PSc000478, PSc010120, PSc010186, PSc020026, PSc030013, PSc052281, PSc110163, PSc130394, PSc130732, PSc350614, PSc390545, PSc400050, PSc480585, and PSc550061. 3 Purity refers to the fraction of a given photometric class that belongs to the equivalent spectroscopic class ). Hosseinzadeh et al. 2020). This selection cut reduces the sample size from 53 to 36 events (Table 1, row 3).

Number of Light Curve Data Points
Both the classification confidence and the ability to meaningfully model the light curves with MOSFiT ( §3.1) are affected by the number of light curve data points; namely, the number of data points relates to the ability to constrain the MOSFiT models and return statistically meaningful posterior distributions. Here we set a threshold of ≥ 11 data points total across the four observed filters (griz) to match the number of model free parameters 4 . This selection cut reduces the sample size from 36 to 24 (Table 1, row 4).

Model Convergence
The aforementioned selection criteria are applied prior to modeling. After all three criteria are applied, we model the 24 photometrically-classified SLSNe with a magnetar central engine model, implemented in MOSFiT. Although we have reduced our sample to identify only events with a sufficient number of data points and high confidence as SLSNe, light curves with marginal detections or potentially misclassified events could in principle survive the above pre-modeling selection metrics. Therefore, we include an additional cut based on the model convergence factor as measured by calculating the Gelman-Rubin statistics, or potential scale reduction factor (PSRF; Gelman & Rubin 1992), which estimates the extent to which the full parameter space has been explored in our MCMC models. Brooks & Gelman (1998) suggests that PSRF < 1.2 provides reliable convergence, but we set a stricter threshold of PSRF < 1.1 as done in Nicholl et al. (2017b) and Hsu et al. (2021), which is also the termination value for our mod-4 One parameter is set to have a constant value, leaving us with 11 free parameters; see §3.1.  els (see §3.1). This post-modeling selection cut reduces our sample size from 24 to 19 (Table 1, row 5). Our final photometric sample consists of 10 events classified as SLSNe by both algorithms, with the remaining 9 classified as SLSN by either SuperRAENN or Superphot. See Table 2 for the predicted SN type of each transient in our final sample and their respective classification confidence.

Justification of Our Choices
In Figure 1 we show the combined effects on the final sample size of varying the minimum classification confidence and the number of data points; we use this as a guide such that our final sample consists of events with sufficient confidence level and data points to obtain a robust model. In each cell we show the number of events that survive each pair of minimum threshold for  confidence and number of detections, and we quote the final sample size after applying both the AGN and convergence cuts in parentheses. To extract a comparable sample size to the PS1-MDS spectroscopic sample (17 events) that will return statistically meaningful results, we outline in Figure 1 the combinations of minimum confidence and detection thresholds that produce a minimal final sample size ≥ 17. We find that our choice of minimum confidence (≥ 0.5) and number of detection (≥ 11) falls within the outlined region, indicating that our selection criteria are reasonable and justified.

Brief Description of the Model
We fit the optical light curves of the 19 photometrically-classified SLSNe (selected as described in §2) using the Modular Open-Source Fitter for Transients (MOSFiT; Guillochon et al. 2018) with the magnetar spin-down model described in Nicholl et al. (2017b).
MOSFiT is an open-source, Python-based light curve fitting package that employs a Markov chain Monte Carlo (MCMC) algorithm to fit a one-zone, grey-opacity ana-
The magnetar model has 12 free parameters, of which 8 are nuisance parameters that we marginalize over to obtain the 4 key physical parameters related to the ejecta and engine properties. We fix one of the nuisance parameters, the angle θ P B between the magnetic field and the rotational axis of the magnetar, to 90 • as this ensures that the derived B-field strength is a lower limit (following Nicholl et al. 2017b). The nuisance parameters, κ, κ γ , M NS , n H,host , are not well-constrained by the model. Events with sufficient late-time observations may constrain the γ-ray opacity κ γ , but this is not the case for our sample. The neutron star mass is degenerate with the spin period and magnetic field strength but is not well constrained. The explosion time, t exp , is the time between explosion and first observation in the pure magnetar model. The main parameters that constrain the observed properties of a SLSN are the neutron star's initial spin period, P , magnetic field strength, B, ejecta mass, M ej , and ejecta velocity, v ej (the latter two can be combined to determine the kinetic energy, E K ). The model parameters and their priors are listed in Table 3.
For each light curve fit, the first 10,000 iterations are used to burn in the ensemble, during which minimization is employed periodically as the ensemble converges to the global optimum; the remainder of the run-time is used to sample the posterior probability distribution. Convergence is measured by calculating the PSRF, and we terminate our fits when PSRF < 1.1. Most events typically require 30,000-60,000 iterations to reach convergence, depending on the number of data points and the scatter around our model.

Light Curve Fits
In Figure 2, we show the magnetar model light curve fits for the 19 PS1-MDS photometrically-classified SLSNe. The shaded regions are the MOSFiT light curve fits, where the upper and lower bounds are 1σ uncertainties calculated from the 120 MCMC walkers, while the solid light curves are based on the parameter medians. To allow for a proper comparison with the PS1-MDS spectroscopic sample, we also show in Figure 3 the 17 PS1-MDS spectroscopically-classified SLSNe (which were used in the classification training samples). Previous studies (Nicholl et al. 2017b;Villar et al. 2018;Blanchard et al. 2020) have already modeled all but one of these events (PS1-12cil) in the same manner as this work. Two peculiar events from the PS1-MDS spectroscopically-classified sample, PS1-11ap and PS1-12cil, exhibit post-peak undulations (e.g., Inserra et al. 2013;Nicholl et al. 2014;Inserra et al. 2017;Nicholl et al. 2016a;Hosseinzadeh et al. 2021). Since our MOSFiT model does not account for these "bumps", we replace the model of PS1-11ap from Blanchard et al. (2020) with the version presented in Hosseinzadeh et al. (2021), which converts these bumps into upper limits prior to fitting. We also include the model for PS1-12cil from Hosseinzadeh et al. (2021) to complete the PS1-MDS spectroscopic SLSN sample. For illustrative purposes, we extrapolate all light curves (both photometric and spectroscopic samples) back to the inferred explosion time and forward 100 days after the last detection.
Overall, we find that the model fits the observed light curves well, and is better constrained for events with more extensive data. The resulting median values and 1σ uncertainties for the four main physical parameters (P , B, M ej , v ej ), calculated based on the posterior probability distributions from 120 MCMC walkers are summarized in Table 4  and spectroscopic samples are listed in Table 5. We also list in Table 5     Note-The median values and 1σ ranges for the magnetar engine and ejecta parameters of the PS1-MDS SLSN samples (photometric and spectroscopic), and the SLSN compilation sample (from Hsu et al. 2021, with the addition of PS1-12cil), which include the 17 PS1-MDS spectroscopicallyclassified SLSNe.

Observational Properties
The PS1-MDS samples (both spectroscopic and photometric) collectively span a wide range of redshifts, z ≈ 0.3 − 2. To properly compare the observational properties of the PS1-MDS SLSNe, we correct their observed peak apparent magnitudes to a single rest-frame filter. Since we do not have a complete set of spectra for the spectroscopic sample, and by definition no spectra for the photometric sample, we do not apply a complete K-correction; instead, we apply only a cosmological K-correction factor of 2.5 log 10 (1+z) to the peak magnitude in the band closest to the rest-frame g-band for each event and correct for Milky Way extinction. We plot the resulting peak g-band absolute magnitudes as a function of redshift in Figure 4. The PS1-MDS spectroscopic sample spans a range of ≈ −20.5 to ≈ −22.6, while the photometric sample spans a wider range of ≈ −18.7 to ≈ −22.6. As expected, lower luminosity SLSNe are restricted to lower redshift (z 0.5), while higher luminosity events are distributed to higher redshift (z ≈ 2). The spectroscopic sample is intrinsically more luminous, with a median peak magnitude of −22 as compared to −20.8 for the photometric sample.
We also plot in Figure 4 the per-visit PS1-MDS limiting magnitude of ≈ 23.3 , as well as the effective spectroscopic follow-up limit of ≈ 22.5 (Lunnan et al. 2018). The majority of the photometric sample have peak absolute magnitudes either around or PSc480628 PSc490019 0 5 10 15 below the spectroscopic follow-up depth, which explains why these event were not chosen for spectroscopic followup. However, there are 5 photometrically-classified SLSNe (PSc061198, PSc080492, PSc110446, PSc390605, and PSc490019) at lower redshift (z ≤ 0.6) that are more than 1 magnitude brighter than the threshold but were not chosen as follow-up candidates.

Physical Properties and Correlations
In Figure 5 we show two-dimensional distributions of the primary physical parameters (P , B, M ej , and v ej ; the medians of the posteriors) and redshifts of both PS1-MDS samples and the SLSN compilation, which contains events from a wide range of surveys (including the PS1-MDS spectroscopic sample). We explore both differences between the three samples, and parameter correlations for the combined sample (all three samples together, 101 SLSNe in total). Specifically, we compare the PS1-MDS photometric sample and the spectroscopic compilation sample using the two-sample Kolmogorov-Smirnov (K-S) test (Smirnov 1948) and the two-sample Anderson-Darling (A-D) test (Anderson & Darling 1952). Both tests are designed to determine whether two distributions arise from the same underlying population. The A-D test is a modification of the K-S test that is more sensitive to the tails of a distribution, whereas the K-S test gives more weight to the mean of a distribution. We report the resulting p-values from these tests, to determine if both are drawn from the same parameter distribution, at the top of each column in Figure 5.
The differences in redshift distributions between the two samples reflect the design characteristics of the various surveys (e.g., PS1-MDS, Dark Energy Survey, PTF, etc). In terms of the magnetar model parameters we find that the distributions are overall in good agreement, except for the ejecta velocity, which has statistically significant p-values for the A-D test. This indicates that we can reject the null hypothesis at 95% conficence that the ejecta velocity for the photometric and the spectroscopic compilation samples are drawn from the same distribution. This may be caused by the sensitivity of the A-D test to tail distributions. The spectroscopic compilation sample spans a range of v ej ≈ (3.6 − 16) × 10 3 km s −1 , while the photometric sample spans a range of v ej ≈ (2.2 − 14) × 10 3 km s −1 , with two events 6 (PSc130096 and PSc390605) having v ej values that fall outside the range of the spectroscopic population. Removing these two outliers return an updated A-D test p-value of ≈ 0.06, suggesting that other than these two specific data points, the remainder of the photometric sample fit into the spectroscopic sample well. We explore the posterior distributions of the magnetar parameters from the photometric sample in more detail in the next subsection.
As done in previous SLSN parameter studies (e.g., Blanchard et al. 2020;Hsu et al. 2021), we combine the PS1-MDS photometric and literature samples to confirm known correlations and explore new ones. For each pair of parameters, we perform a Monte Carlo procedure to calculate the Spearman rank correlation coefficient (ρ; Spearman 1904) and its associated 1σ bound using the method described in Curran (2014). The results are summarized in each panel of Figure 5. We find the same results as Hsu et al. (2021), where most parameter combinations exhibit either no correlation, mild correlations, or mild correlations that are primarily due to the absence of events in specific areas of the parame-K-S Test p = 0.02 A-D Test p = 0.02  Figure 4). The models for PS1-11ap and PS1-12cil are both obtained from Hosseinzadeh et al. (2021). The gray crosses mark the remaining spectroscopically confirmed SLSNe from Hsu et al. (2021). In the top panels we show the parameter distributions for the PS1-MDS photometric sample (blue), PS1-MDS spectrosc sample (red), and the SLSN compilation sample (grey), along with the median p-values associated with both the K-S test and the A-D test statistics, calculated using the PS1-MDS photometric sample and the SLSN compilation. In each panel we quote the median value and 1σ bound of the Spearman rank correlation coefficient using the PS1-MDS photometric sample and literature data set. Of all parameter pairs, P and Mej exhibit the strongest correlation, consistent with the findings in Blanchard et al. (2020) and Hsu et al. (2021).
ter space. The mass-spin correlation discussed first in Blanchard et al. (2020) remains strong after merging the photometric and spectroscopic samples. All other mild correlations have been previously explained as being due to observational biases in Blanchard et al. (2020) and Hsu et al. (2021), and we do not find any new statistically significant correlations here.

Posterior Distributions of the Photometric Sample
To explore any differences in magnetar and ejecta parameters between the PS1-MDS photometric and spectroscopic samples, we show in Figure 6 the joint posterior distributions of the PS1-MDS photometric, PS1-MDS spectroscopic, and the compilation samples. We construct the joint posterior distributions by selecting 100 randomly sampled walkers from each MOSFiT fit. 7 To capture uncertainties in the test statistics, we calculate and report in each panel the two-sample K-S test and the two-sample A-D test p-values between the PS1-MDS photometric sample and the spectroscopic compilation using a modified bootstrap method. For each parameter, we calculate a distribution of p-values by repeating the following procedure 5000 times. We assemble a joint posterior for the 19 photometrically-classified SLSNe by randomly drawing one MCMC walker from the individual posterior for each event, and we do the same for the 82 spectroscopically-classified SLSNe. We then calculate p-values for the K-S and A-D tests comparing these two joint posteriors. We report the median and 1σ bounds of these distributions of the resulting p-values on top of each panel in Fig 6. The posterior distributions for the physical parameters are in good agreement, except for v ej , as noted previously; removing PSc130096 and PSc390605 from the photometric sample leads to p = 0.11 +0.15 −0.08 (K-S) and p = 0.05 +0.06 −0.03 (A-D). We also note that while the K-S and A-D tests indicate that the distributions of P and M ej are drawn from the same distribution, the photometric sample skews to slower spins and lower ejecta masses (this trend is still in agreement with the massspin correlation). This difference can be ascribed to the systematically lower luminosities of the photometric SLSNe ( Figure 4) compared to the PS1-MDS spectroscopic SLSNe.

Effects of Classification Uncertainty
As indicated in Table 2, 9 of the 19 photometricallyclassified SLSNe in our final sample were designated as 7 We take 100 here instead of the full 120 walkers as described in §3.2 because some events modeled previously in Blanchard et al. (2020) only have 100 walkers.
SLSNe by only one of the two classifiers. To investigate the impact of these cases of classification disagreement, we repeat the analyses in the previous subsections using only events classified as SLSNe by both Superphot and SuperRAENN. This "consensus" photometric sample spans a peak absolute magnitude range of ≈ −20.3 to ≈ −22.6. However, despite excluding some of the lowest luminosity events, the median peak magnitude is still ≈ 1 mag dimmer than that of the spectroscopic sample (see Figure 7, left), and we find the same trend of systematically lower luminosity at any redshift as seen for the full sample in Figure 4. Our conclusion about the lower luminosities probed by the photometric sample thus remains unchanged. Systematically removing objects classified as SLSNe by only one classifier eliminates the disagreement in the v ej distributions but introduces mildly statistically significant differences in B and M ej . The consensus sample shifts to higher ranges of B ≈ (1 − 7.7) × 10 14 G, v ej ≈ (0.37−1.41)×10 4 ) km s −1 , a lower range of M ej ≈ 1.4−9.9 M , and a similar range of P ≈ (1.17−7.98) ms in parameter distributions. These shifts are all consistent and expected for SLSNe with higher luminosities. See Figure 7 for these changes in magnetar model parameters. The shift in B is reflected in the posterior distribution but not as strongly in M ej .

DISCUSSION AND CONCLUSIONS
In this paper we presented a case study for timedomain science with machine learning-based photometric classification, focusing on SLSNe from the PS1-MDS. Our analysis consisted of two critical aspects that would need to be undertaken for any future such studies (for SLSNe or any other types of transients). First, we began with a sample of events nominally classified as SLSNe by two independent machine learning-based pipelines (SuperRAENN and SuperPhot). We then applied various selection criteria to increase the sample purity (e.g., removing likely AGN flares, setting a higher minimum classification probability threshold) at the cost of sample completeness. Our sample size following these cuts was 36% of the initial sample (24 of the 67). Subsequent to the sample refinement we carried out modeling with MOSFiT to extract physical parameters in order to compare the photometric sample with existing spectroscopic samples modeled in the same way. The requirement for model convergence eliminated 5 additional events from the sample (21% reduction from 24 to 19). These two critical steps of sample refinement and modeling will be essential for all studies with photometrically-classified samples.  Comparing our photometric SLSN sample to the PS1-MDS spectroscopically-classified SLSNe and to the larger sample of spectroscopic SLSNe, we find an overall similarity in both observed properties and inferred magnetar and ejecta parameters. We do note a potential shift in the photometric sample to slower magnetar spins and lower ejecta masses, which may reflect the fact that the photometric SLSNe are systematically dimmer than the spectroscopic PS1-MDS SLSNe (due to the shallower effective magnitude limit required for spectroscopy). If this is indeed the case, then it highlights an important advantage of photometric classification in deep surveys (such as PS1-MDS and LSST).
Our initial classifications and the subsequent modeling both rely on the existence of redshift information. In the case of our PS1-MDS sample, the redshifts were determined from host galaxy spectroscopy after the survey concluded. Such data may be difficult to obtain for the large samples expected from LSST (e.g., 10 6 SNe per year, and ∼ 10 4 SLSNe per year Villar et al. 2018). However, robust photometric redshifts are likely to be as useful as spectroscopic redshifts. We also note that one source of contamination in our initial photometric sample appears to be AGN (21%, 14 of 67 events) despite the fact that the PS1-MDS sample was designed to eliminate variable AGN. These contaminating AGN were again identified via host galaxy spectroscopy, which will not be available for the LSST samples; a more robust elimination of AGN will be essential.
Overall, our analysis highlights some challenges in constructing pure samples of photometrically-classified SNe, but we believe that these challenges are surmountable. The photometric sample explored here is smaller than the overall known spectroscopic sample by a factor of several, but looking forward to LSST, even a highly conservative selection with relatively low completeness will easily exceed the spectroscopic sample by two orders of magnitude.
The Berger Time Domain group at Harvard is supported in part by NSF and NASA grants, including support by the NSF under grant AST-2108531, as well as by the NSF under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions http://iafi.org/). VAV acknowledges support in part by the NSF through grant AST-2108676. The computations presented in this work were performed on the FASRC Cannon cluster supported by the FAS Division of Science Research Computing Group at Harvard University. for the full photometric sample, but is still ≈ 1 magnitude dimmer than the median of ≈ −22 for the spectroscopic sample. Even though the full and the consensus photometric samples have comparable median B values (≈ 1.74 × 10 15 G and 1.80 × 10 15 G for the full and consensus samples, respectively), the consensus sample spans a much narrower and higher range in B. The shift in Mej is more strongly reflected, with a lower median value (≈ 3.58 M , full; ≈ 2.33 M , consensus) at a lower range. All of the shifts in magentar model parameters are consistent with SLSNe with higher luminosities than the full photometric sample.