reLAISS: Python Tool for Supernova Similarity

• reLAISS is a Python library that enables similarity searches and anomaly prioritization by leveraging ZTF light curves and Pan-STARRS host-galaxy photometry.
• It integrates interpretable light-curve features with host context using Monte Carlo uncertainty-aware retrieval and ANN indexing to enhance comparative analysis.
• The framework replaces random-forest anomaly scoring with a Generalized Pareto distribution fit for calibrated, real-time anomaly detection.

Searching arXiv for the specified paper to ground the article in the source preprint. reLAISS is a Python library for supernova similarity search and anomaly prioritization that was introduced as “a modified framework for similarity searches of supernovae using extracted features of ZTF light curves and Pan-STARRS host galaxy photometry” and is “built on the original LAISS framework” (Reynolds et al., 17 Jul 2025). It is designed for the time-domain astronomy problem of determining which previously observed supernovae and host galaxies are most similar to a newly discovered transient, and whether the object is unusual enough to merit rapid follow-up. The framework is motivated by survey conditions in which discovery rates are expected to exceed one million supernovae per year and only about 1% of events can realistically receive spectroscopic follow-up, making photometric classification alone insufficient for isolating rare or scientifically unusual objects (Reynolds et al., 17 Jul 2025).

1. Scientific setting and motivation

reLAISS was developed for a data-deluge regime in time-domain astronomy in which the Vera C. Rubin Observatory is expected to discover more than one million supernovae per year, while only about 1% of events can realistically receive spectroscopic follow-up (Reynolds et al., 17 Jul 2025). Within that setting, the operational problem is not merely taxonomic classification, but triage: finding rare or outlier transients among large numbers of routine supernovae, matching a transient to the most relevant historical analogs, incorporating host-galaxy context as a physical prior on progenitor environment, and ranking events for rapid follow-up.

The framework is explicitly oriented toward “one-in-a-million” discovery. It addresses cases in which a transient may be unusual within an existing class, may exhibit an atypical combination of light-curve behavior and host environment, or may be scientifically valuable even when sparse data preclude confident formal classification. The emphasis on host context reflects the paper’s statement that host galaxies constrain the likely stellar population when transient data are sparse, which is valuable for comparative studies and for identifying unusual events embedded in particular environments (Reynolds et al., 17 Jul 2025).

A central conceptual feature of reLAISS is that it is not just a classifier. It is a retrieval and ranking system that supports the question “What is this most like?” rather than forcing every event into a rigid taxonomy. This suggests a methodological shift from closed-label assignment toward similarity-based comparative analysis, especially for follow-up prioritization under extreme alert volumes.

2. Relation to LAISS and principal design changes

reLAISS is described as “built on the original LAISS framework,” where LAISS denotes “Light curve Anomaly Identification and Similarity Search,” but with several major modifications (Reynolds et al., 17 Jul 2025). The revised system expands the earlier framework in multiple directions.

First, reLAISS replaces the prior “light-curve-calculated features” with a new light-curve feature set derived directly from ZTF light curves and described as more interpretable (Reynolds et al., 17 Jul 2025). Second, host-galaxy photometry is integrated more explicitly: reLAISS couples interpretable light-curve morphology features with extinction-corrected Pan-STARRS host-galaxy photometric features so that explosion physics and associated stellar populations can be probed simultaneously (Reynolds et al., 17 Jul 2025). Third, the host-association method changes from GHOST to Prost, which the paper describes as replacing GHOST (Reynolds et al., 17 Jul 2025).

Additional changes concern flexibility and uncertainty handling. Users can query the system using light-curve features only, host-galaxy features only, or both together; when both are used, a weighting term allows one domain to be prioritized relative to the other (Reynolds et al., 17 Jul 2025). The framework also incorporates uncertainty-aware retrieval, including Monte Carlo resampling when queried features are uncertain, so that neighbors can be re-queried and re-ranked when features are poorly constrained (Reynolds et al., 17 Jul 2025).

The anomaly-detection component is also changed. LAISS used a random forest for real-time anomaly detection, whereas reLAISS replaces this with a Generalized Pareto distribution fit to the upper tail of neighbor distances, followed by a sigmoid transformation that maps anomaly probabilities to a 0–100 score (Reynolds et al., 17 Jul 2025). This suggests an effort to make anomaly calibration more directly tied to the geometry of nearest-neighbor distances in feature space.

3. Observational inputs and feature representation

reLAISS uses two principal observational data sources: ZTF alert photometry, from the ANTARES broker or user-provided photometry files, in the ZTF $g$ and ZTF $r$ bands; and the Pan-STARRS 3$\pi$ host-galaxy catalog, with hosts identified using Prost and host photometry measured in $griz$ (Reynolds et al., 17 Jul 2025).

For each light-curve band, the framework extracts interpretable morphology features including peak magnitude, time from first observation to peak, rise time to half-flux, decline time to half-flux, duration above half-flux, and rolling variance (Reynolds et al., 17 Jul 2025). It also computes color-based quantities: $g-r$ color at peak, mean $g-r$ across all detections, and mean change in $g-r$. These color quantities are obtained using linear interpolation to align observations to a common phase (Reynolds et al., 17 Jul 2025).

A further light-curve descriptor is the median local curvature of the photometry in each band, evaluated in the 20 days before peak and the 20 days after peak. The curvature is estimated using the centered second finite difference of the magnitudes:

$$\kappa_i = \frac{m_{i+1} - 2m_i + m_{i-1}}{\left[\frac{1}{2}(t_{i+1} - t_{i-1})\right]^2}$$

This is presented as a compact way to quantify light-curve shape and sharpness around peak (Reynolds et al., 17 Jul 2025). The paper emphasizes that the full feature set provides an “optimal balance between interpretability and flexibility,” meaning that the descriptors remain physically readable while still serving as an effective basis for similarity search (Reynolds et al., 17 Jul 2025).

After host association, reLAISS computes host-galaxy photometric properties from Pan-STARRS: extinction-corrected Kron magnitude in $g,r,i,z$, together with host-galaxy colors $g-r$, $r$0, and $r$1 (Reynolds et al., 17 Jul 2025). These colors are independently weighted so that outliers caused by catastrophic color-estimation failures do not dominate the similarity search (Reynolds et al., 17 Jul 2025). The extinction correction is explicitly noted as important because raw observed colors can be biased by line-of-sight dust (Reynolds et al., 17 Jul 2025).

4. Host association and similarity-search procedure

Host association in reLAISS is carried out with Prost. The method estimates the posterior probability of association using a Monte Carlo simulation over observed galaxy properties together with user-defined priors on the true host’s redshift, fractional offset, and absolute magnitude (Reynolds et al., 17 Jul 2025). This host-association stage is a major methodological shift relative to LAISS’s use of GHOST (Reynolds et al., 17 Jul 2025).

For similarity retrieval, reLAISS uses ANNOY to build an indexed reference space and perform approximate nearest-neighbor search through locality-sensitive hashing (Reynolds et al., 17 Jul 2025). The search can be performed in three modes: light-curve features only, host features only, or both (Reynolds et al., 17 Jul 2025). When both are included, a user-defined weighting term determines how much each feature set contributes to the search (Reynolds et al., 17 Jul 2025).

The user may specify the number of neighbors $r$2. If $r$3 is not specified, reLAISS uses the kneed package to identify the “knee” in the distance-versus-neighbor curve and suggest a reasonable number of relevant neighbors (Reynolds et al., 17 Jul 2025). This suggests a pragmatic mechanism for adapting neighborhood size to local structure in the reference space rather than imposing a globally fixed $r$4.

If feature uncertainties are available, reLAISS can re-query and re-rank neighbors using Monte Carlo perturbations of the input features (Reynolds et al., 17 Jul 2025). The paper does not provide a closed-form equation for this uncertainty propagation, but it describes the mechanism qualitatively: uncertain input features are sampled and the search is repeated in order to stabilize the retrieved neighbors (Reynolds et al., 17 Jul 2025). A plausible implication is that the framework is intended to remain useful even for early, noisy, or incompletely constrained transients.

5. Anomaly scoring and calibration

reLAISS includes a new real-time anomaly score calibration. Instead of the random-forest anomaly detector used in LAISS, it fits a Generalized Pareto distribution to the upper tail of the nearest-neighbor distance distribution from the reference database (Reynolds et al., 17 Jul 2025). The resulting real-time feature probabilities for the queried transient are then mapped by a sigmoid transformation into an anomaly score on a 0–100 scale (Reynolds et al., 17 Jul 2025).

This anomaly score is described as calibrated and interpretable, with the purpose of flagging unusual transients (Reynolds et al., 17 Jul 2025). Because the score is derived from the upper tail of neighbor distances, it is tied to how far the query lies from its nearest analogs in the indexed feature space. This suggests a distance-based notion of unusualness rather than a supervised novelty label.

The anomaly framework is demonstrated in the paper through a long-duration SN II, ZTF20acpjbgc, whose light curve and evolving anomaly score over time are shown as a proof of concept for the real-time anomaly detector (Reynolds et al., 17 Jul 2025). The use of evolving scores over time indicates that anomaly assessment is intended to be dynamic as additional photometric information becomes available.

6. Software package, configurability, and demonstrations

reLAISS is released as a pip-installable Python package with an accompanying reference set of 20,000 archival light-curve and host-galaxy feature vectors, a set of tutorials, and source code on GitHub (Reynolds et al., 17 Jul 2025). The reference bank consists of archived transients with extracted ZTF light-curve features and matched Pan-STARRS host-galaxy features (Reynolds et al., 17 Jul 2025). The packaging is significant because it makes the method reproducible, easy to adopt, usable in automated pipelines, and transparent and extensible (Reynolds et al., 17 Jul 2025).

The user-facing configuration options explicitly mentioned in the paper include the number of neighbors returned, feature-set selection, host/light-curve weighting, uncertainty-aware retrieval through Monte Carlo simulations, optional manual host specification for theoretical models, and PCA/dimensionality options (Reynolds et al., 17 Jul 2025). The figure caption notes a demonstration with 20 Monte Carlo simulations without PCA, implying that PCA is configurable or optional in the workflow, though the paper does not elaborate its full behavior (Reynolds et al., 17 Jul 2025).

The package is aimed at transient astronomers, supernova researchers, survey pipeline scientists, follow-up coordinators, and others building tools for large-scale transient triage (Reynolds et al., 17 Jul 2025). Representative use cases include finding nearest historical analogs to a new supernova, prioritizing spectroscopy, identifying peculiar or extreme events, studying the relation between host environment and transient class, comparing observed events to theoretical models, and rapidly flagging anomalies in survey streams (Reynolds et al., 17 Jul 2025).

The paper includes two demonstration examples. In one, panels show light-curve neighbors and host-galaxy neighbors for ZTF23aagpjyp, identified as an SN Ia-CSM, using light-curve-only and host-only retrieval modes, respectively; these searches used 20 Monte Carlo simulations without PCA (Reynolds et al., 17 Jul 2025). In the other, the anomalous transient ZTF20acpjbgc is used to illustrate the anomaly detector (Reynolds et al., 17 Jul 2025). These examples function primarily as workflow demonstrations rather than as a full benchmark study.

7. Interpretation, scope, and limitations

The stated scientific value of reLAISS lies in triaging rare and interesting events that future surveys will uncover by combining ZTF light-curve morphology, Pan-STARRS host-galaxy photometry, approximate nearest-neighbor retrieval, uncertainty-aware ranking, and real-time anomaly scoring (Reynolds et al., 17 Jul 2025). Its operating logic is comparative rather than purely classificatory: it organizes the discovery stream by similarity and unusualness while retaining interpretable descriptors tied to light-curve shape and host stellar population context.

A common misconception would be to treat reLAISS as a replacement for formal photometric classification. The paper instead presents it as a tool for similarity search and anomaly prioritization, particularly useful when photometric classification alone will not be sufficient to isolate rare events or prioritize the small fraction of transients that can receive spectroscopic follow-up (Reynolds et al., 17 Jul 2025). Another possible misconception is that the system uses only transient photometry; in fact, a defining feature of reLAISS is its explicit coupling of transient and host-galaxy information (Reynolds et al., 17 Jul 2025).

The paper is brief and “mostly presents the package and workflow rather than a detailed benchmark study” (Reynolds et al., 17 Jul 2025). Accordingly, it does not provide an extensive comparative evaluation across large labeled benchmarks, nor does it give a closed-form equation for uncertainty propagation in Monte Carlo re-querying (Reynolds et al., 17 Jul 2025). This suggests that the main contribution is infrastructural and methodological: a survey-ready similarity-search framework with interpretable feature engineering, flexible query modes, and anomaly calibration for high-volume transient operations.

In that sense, reLAISS occupies a specific niche in Rubin-era transient astronomy. It provides a practical mechanism for asking which historical events and environments most resemble a new detection, while simultaneously estimating whether the event is sufficiently atypical to justify rapid follow-up in an observing regime where most discoveries cannot be examined spectroscopically (Reynolds et al., 17 Jul 2025).