Superphot+: Photometric SN Pipeline

• The paper introduces Superphot+ as a photometric supernova classification pipeline that extracts compact features from an empirical parametric light-curve model.
• It employs joint multiband fitting and Bayesian posterior-based priors to enhance classification accuracy on sparse, irregular survey photometry.
• The pipeline achieves competitive F1-scores and accuracy in redshift-independent mode, making it effective for realtime ZTF and Rubin/LSST transient discovery.

Superphot+ is a photometric supernova-classification pipeline that fits multiband light curves with an empirical parametric model, converts the fitted parameters into compact features, and classifies the result with a supervised machine-learning model. It was introduced as a redshift-agnostic, realtime-capable successor to Superphot, with explicit deployment on survey alert streams and design goals aligned with ZTF and Rubin/LSST-scale transient discovery (Soto et al., 2024).

1. Definition and scientific role

Superphot+ addresses the central survey-era problem that photometric discovery rates vastly exceed spectroscopic follow-up capacity. Its classification task is multiclass rather than binary, with five output labels: SN Ia, SN II, SN Ib/c, SN IIn, and SLSN-I. In its default mode it does not rely on redshift information, and the motivating claim of the method is that it can still maintain performance comparable to redshift-dependent classifiers while remaining suitable for realtime operation (Soto et al., 2024).

Methodologically, Superphot+ is a hybrid system. It is not a direct template matcher in the classical spectro-photometric sense, and it is not a pure end-to-end neural sequence model. Instead, it combines model-based feature extraction with supervised classification: a parametric light-curve model is fit first, and the resulting fit parameters are then passed to a gradient-boosted machine. This design makes the classifier interpretable at the feature level, because the dominant inputs correspond to rise timescales, fall timescales, plateau durations, peak band ratios, and related phenomenological descriptors of SN evolution (Soto et al., 2024).

The intended operational regime is large, heterogeneous survey photometry rather than curated, fully sampled light curves. This is reflected in three explicit design choices: it is trained on real ZTF data rather than simulations alone, it supports partial light curves, and it is integrated into broker infrastructure for realtime use. The same architecture was subsequently applied to Rubin Data Preview 1 commissioning data, where it served as the terminal subtype classifier after candidate discovery and vetting, rather than as the discovery engine for the full alert stream (Freeburn et al., 30 Jul 2025).

2. Parametric light-curve model and multiband representation

The empirical light-curve model used by Superphot+ is inherited from the Villar et al. formulation and is written as

$F(t) = \frac{A}{1+\exp\!\left(-\frac{t-t_0}{\tau_{\mathrm{rise}}}\right)} \begin{cases} 1-\beta (t-t_0), & t-t_0 < \gamma,\[4pt] (1-\beta\gamma)\exp\!\left(\frac{\gamma-(t-t_0)}{\tau_{\mathrm{fall}}}\right), & t-t_0 \ge \gamma. \end{cases}$

Each band has seven parameters: $A$, $t_0$, $\tau_{\mathrm{rise}}$, $\beta$, $\gamma$, $\tau_{\mathrm{fall}}$, and $\sigma_{\mathrm{extra}}$. The model accounts for a rise in brightness, followed by an optional plateau and subsequent decline, which is why it can represent both thermonuclear and core-collapse morphologies at a coarse phenomenological level (Soto et al., 2024).

A major methodological change relative to earlier Superphot is joint multiband fitting. Rather than fitting $g$ and $r$ independently, Superphot+ fits them simultaneously in a 14-dimensional space, with the $A$0-band treated as the reference band. For all parameters except $A$1, the sampled $A$2-band quantity is the logarithm of the ratio of the $A$3-band parameter to the corresponding $A$4-band parameter; for phase, the sampled quantity is the difference $A$5. This introduces explicit cross-band correlation and is meant to stabilize inference when one band is sparsely sampled (Soto et al., 2024).

The Bayesian formulation is likewise explicit. The posterior is

$A$6

with individual observations modeled as

$A$7

The $A$8 term is added in quadrature to absorb empirical-model mismatch. The paper also imposes shape constraints to ensure that the piecewise transition occurs after a plateau-like region and that the second segment declines more steeply than the first (Soto et al., 2024).

This representation is deliberately survey-native rather than rest-frame normalized. Photometry is converted from magnitude to flux with $A$9 using $t_0$0, Milky Way extinction is corrected using $t_0$1, the Fitzpatrick and Massa extinction law, and $t_0$2, and the default redshift-free mode leaves the light curves in the observer frame. Host extinction is not corrected, and when redshift is later included it is appended as a feature rather than used to transform the light curves themselves (Soto et al., 2024).

3. Training set, fitting strategy, and classifier construction

The spectroscopic training and evaluation sample comprises 6,061 ZTF supernovae that pass data-quality cuts and fit-quality filtering. The class distribution is strongly imbalanced: 4,546 SN Ia, 978 SN II, 259 SN Ib/c, 257 SN IIn, and 83 SLSN-I. The source sample was assembled from ZTF light curves obtained through October 2023, with spectroscopic labels from TNS (Soto et al., 2024).

Input pruning is an important part of the pipeline. A light curve must have at least five points with $t_0$3 in each of $t_0$4 and $t_0$5. In both bands, the maximum amplitude must exceed $t_0$6 the band’s mean flux uncertainty, and the flux standard deviation must exceed the band’s mean flux uncertainty. The pipeline also clips late-time subtraction residuals in $t_0$7 and $t_0$8 rather than introducing a constant-offset parameter; this affected at least one point in 4,324 of 9,526 spectroscopic light curves and fewer in the photometric sample (Soto et al., 2024).

The priors are learned iteratively from the fitted ZTF population. The procedure starts from broad uniform priors, fits the dataset, combines posterior marginals across light curves, rebalances by class before combining, replaces the priors with those population-level marginals, and repeats until priors and posteriors are similar. The final priors are truncated Gaussians or truncated log-Gaussians. This iterative prior-learning scheme is one of the stated reasons for the improvement over earlier Superphot (Soto et al., 2024).

Superphot+ evaluates several fitters. Nested sampling is used for archival and training-set fitting; SVI in NumPyro/JAX is used in realtime; and in ANTARES an SVI fit is replaced by nested sampling if the SVI result is poor. The reported nested-sampling configuration uses dynesty with 50 live points, 5,000 maximum iterations, stopping criterion $t_0$9, a one-ellipsoid unimodal assumption, and a random-walk proposal for new live points. SVI is much faster but tends to underestimate variance (Soto et al., 2024).

A modified reduced chi-squared is used as the fit-quality filter:

$\tau_{\mathrm{rise}}$0

with $\tau_{\mathrm{rise}}$1, and only light curves with $\tau_{\mathrm{rise}}$2 are retained. This removed 62 spectroscopic objects from 6,123 and 415 photometric objects from 3,973 (Soto et al., 2024).

For classification proper, the default redshift-independent model uses 12 of the 14 fitted parameters, excluding $\tau_{\mathrm{rise}}$3 and $\tau_{\mathrm{rise}}$4. The redshift-inclusive variant adds $\tau_{\mathrm{rise}}$5, redshift $\tau_{\mathrm{rise}}$6, and $\tau_{\mathrm{rise}}$7-corrected $\tau_{\mathrm{rise}}$8-band peak absolute magnitude $\tau_{\mathrm{rise}}$9. The classifier itself is LightGBM with DART boosting, goss sampling, maximum depth 5, maximum 20 leaves, regularization 5, and 250 estimators. To address the extreme class imbalance, Superphot+ uses Bayesian oversampling: posterior draws from minority-class light curves are treated as additional training samples, with 22,730 oversampled feature sets per class across training and validation sets (Soto et al., 2024).

4. Performance, calibration, and early-time behavior

On the five-class spectroscopic dataset, the default no-redshift classifier yields a class-averaged $\beta$0-score of $\beta$1 and a total accuracy of $\beta$2. Including redshift information improves these metrics to $\beta$3 and $\beta$4, respectively. The no-redshift mode also reports class-averaged completeness of 0.73 and class-averaged purity of 0.58 (Soto et al., 2024).

Class-dependent behavior is highly asymmetric. SN Ia is the easiest class, with completeness 0.87 and purity 0.97 in the no-redshift setting. SN II has relatively good purity at 0.84. SN IIn is the hardest class by completeness at 0.52, while SLSN-I and SN Ib/c have low purity because contamination from more numerous classes dominates their predicted samples. A concrete example given by the paper is that only 9% of true SN Ia are mislabeled as SN Ib/c, yet this still produces roughly 420 contaminants, which is over half of the 713 predicted SN Ib/c objects (Soto et al., 2024).

The most important confusion modes are SN Ia versus SN Ib/c, SLSN-I versus SN IIn, and SN II spread across other labels when the plateau is not constrained. The paper attributes these confusions to specific light-curve pathologies: reddened or host-extincted SN Ia can resemble Ib/c, the empirical model cannot capture the SN Ia secondary $\beta$5-band bump well, slow declines in SLSN-I can be fit as plateaus, and sparse or partial sampling can erase the discriminative plateau structure of SN II (Soto et al., 2024).

Superphot+ outputs five-element vectors that sum to one, but the paper explicitly describes them as pseudo-probabilities rather than calibrated probabilities. Calibration analysis indicates that SN Ib/c probabilities are overconfident, SN Ia probabilities are underconfident, and the other classes are fairly well calibrated. The assigned label is always the argmax class, and the associated confidence is the maximum pseudo-probability (Soto et al., 2024).

A high-confidence subset can be extracted by requiring maximum pseudo-probability $\beta$6. This retains 3,200 of 6,061 events, or 52.7%, and raises performance to $\beta$7, accuracy $\beta$8, class-averaged purity $\beta$9, and class-averaged completeness $\gamma$0. The tradeoff is severe for rare classes: 75% of SLSN-I and 75% of SN IIn are removed, whereas only 43% of SN Ia are removed (Soto et al., 2024).

The method also includes an explicit early-phase mode. Because $\gamma$1, $\gamma$2, and their cross-band ratio parameters are poorly constrained in rising-only light curves, an early-phase classifier is trained without them. In truncation experiments, this early-phase classifier outperforms the full-phase classifier before approximately 20 days after peak. Near phase 0, the early-phase model improves SN IIn purity from 0.03 to 0.50, SN Ia completeness from 0.45 to 0.79, and SN II completeness from 0.15 to 0.31 (Soto et al., 2024).

5. Realtime deployment and Rubin DP1 application

Superphot+ is currently integrated as an ANTARES filter and is classifying ZTF alert-stream supernovae in real time. Beyond broker deployment, it was also applied to 3,558 ZTF transients that show SN-like characteristics but lack spectroscopic classifications. In that photometric sample it predicted 58.6% SN Ia, 13.3% SN II, 7.4% SN IIn, 8.0% SLSN-I, and 12.7% SN Ib/c; after bias correction using the purity matrix, the class fractions become 68.5% SN Ia, 17.7% SN II, 6.5% SN IIn, 2.7% SLSN-I, and 4.7% SN Ib/c. Agreement with the ALeRCE light-curve classifier is 82% on the spectroscopic set and 72% on the unlabeled photometric set (Soto et al., 2024).

The Rubin Data Preview 1 study provides the first commissioning-era application to Rubin/ComCam data. There, Superphot+ is not used to mine the full alert stream directly. The pipeline first queries 369,644 DIA objects in three extragalactic fields, filters contaminants through a sequence of automated cuts, reduces the set to 965 candidates for human inspection, and retains 11 likely extragalactic transients. Superphot+ is then applied only at the final subtype-labeling stage (Freeburn et al., 30 Jul 2025).

In that DP1 analysis, the inputs to Superphot+ are flux-offset-corrected, subtracted multi-band light curves built from epoch-binned DIA forced PSF photometry. The model is fit jointly to all available Rubin bands, but only the ComCam $\gamma$3- and $\gamma$4-band fit parameters are used for classification, because the classifier is trained on spectroscopically confirmed ZTF $\gamma$5- and $\gamma$6-band supernovae. The paper describes the classifier as redshift-independent, and for rising-only cases it excludes the parameters that constrain late-time evolution rather than imposing a hard phase-coverage cut (Freeburn et al., 30 Jul 2025).

The DP1 output labels the 11 transients as 6 SN Ia, 2 SN II, 2 SN Ibc, 1 SN IIn, and 0 SLSN-I. The study emphasizes, however, that these are argmax labels from uncalibrated pseudo-probabilities rather than spectroscopic confirmations. Several cases are only weakly separated; for example, 2024aigv is nearly tied among Ia, IIn, and Ibc, and 2024ahyy is only mildly favored as Ia over II and Ibc. Two alternate flux-offset estimates for 2024ahyy and 2024ahzc leave their class assignments unchanged, which functions as a limited robustness check, but the paper is explicit that this is not a full performance-validation study for Rubin data (Freeburn et al., 30 Jul 2025).

A plausible implication of the DP1 application is that Superphot+ is already operationally useful in Rubin-like settings, but that survey transfer remains a domain-shift problem. The paper itself does not report Rubin-native retraining, passband recalibration, or calibrated Rubin class probabilities; instead, it closes by indicating future classification efforts that combine images and photometry (Freeburn et al., 30 Jul 2025).

6. Relation to Superphot, strengths, and limitations

Superphot+ is best understood as an evolution of Superphot rather than a mere rebranding. The earlier Superphot pipeline fit analytical light curves, transformed them into PCA-based features plus peak absolute magnitude, and classified them with a random forest; it depended strongly on redshift because absolute magnitude was a central feature, and in its PS1 implementation it used $\gamma$7 photometry with a two-iteration cross-filter fitting scheme (Hosseinzadeh et al., 2020). Superphot+ replaces that architecture with joint multiband fitting, population-learned correlated priors, Bayesian posterior-based oversampling, and a LightGBM classifier designed to operate even without redshift (Soto et al., 2024).

Aspect	Superphot	Superphot+
Core features	PCA coefficients plus peak absolute magnitude	Fitted parameters from a general parametric SN model
Classifier	Random forest	Gradient-boosted machine using LightGBM
Redshift use	Depends strongly on redshift	Can run without redshift information
Band handling	Two-iteration fitting with weak cross-filter regularization	Joint $\gamma$8+$\gamma$9 fitting in a 14-dimensional space
Operational target	PS1 photometric classification	Realtime ZTF and Rubin-facing classification

This progression also clarifies a common misconception. The PS1-MDS SLSN case study that used Superphot and SuperRAENN jointly did not define an algorithm explicitly named “Superphot+”. What it described was a broader workflow in which Superphot outputs were combined with another classifier, purity and data-quality cuts, and downstream physical modeling with MOSFiT to construct a photometric SLSN sample (Hsu et al., 2022). By contrast, Superphot+ is a distinct classifier paper with its own fitting, feature, and training machinery (Soto et al., 2024).

The principal strengths of Superphot+ are therefore specific and technical: it is redshift-agnostic by design, it is fast enough for realtime operation through SVI-based fitting, it is more interpretable than latent-space deep sequence models because the classifier consumes explicit light-curve parameters, and it is competitive with or better than several prior baselines. On identical ZTF training and evaluation, Superphot+ achieves accuracy 0.83 and $\tau_{\mathrm{fall}}$0 0.61 without redshift, and 0.88 and 0.71 with redshift, compared with Superphot at 0.80 and 0.55. In a four-class redshift-free comparison on the shared subset, it reaches $\tau_{\mathrm{fall}}$1 versus 0.62 for ALeRCE-SN (Soto et al., 2024).

Its limitations are equally explicit. The spectroscopic training set is follow-up selected and not representative of the full transient population; host extinction is not corrected; the empirical model cannot capture secondary SN Ia $\tau_{\mathrm{fall}}$2-band bumps, multiple peaks, or complex late-time behavior well; the pseudo-probabilities are not calibrated; and partial light curves can become prior-dominated. The paper also notes that ignoring time dilation is acceptable at ZTF redshifts but may become more problematic for Rubin, and it suggests redshift-invariant combinations such as $\tau_{\mathrm{fall}}$3, $\tau_{\mathrm{fall}}$4, and $\tau_{\mathrm{fall}}$5 as one possible modification for deeper surveys (Soto et al., 2024).

In sum, Superphot+ occupies a specific methodological niche in transient astronomy: an interpretable, Bayesian-fit-plus-GBM classifier for sparse, irregular survey light curves, optimized for broker deployment and for the regime in which redshift may be missing or delayed. Its later Rubin DP1 use confirms that the framework is portable enough for early Rubin data products, while also making clear that calibration, domain adaptation, and spectroscopic anchoring remain central requirements for survey-scale scientific use (Freeburn et al., 30 Jul 2025).