pith. sign in

arxiv: 2607.00228 · v1 · pith:DZOBLYDFnew · submitted 2026-06-30 · 🌌 astro-ph.IM · astro-ph.HE· cs.LG

Leveraging Multimodality for Real-Time Classification of Transients and Variables found by the Zwicky Transient Facility

Ved G. Shah , Nabeel Rehemtulla , Adam A. Miller , Sushant Sharma Chaudhary , Michael W. Coughlin , Antoine Le Calloch , Matthew J. Graham , Joahan Castaneda Jaimes
show 8 more authors
Theophile Jegou du Laz Ashish A. Mahabal Frank J. Masci Josiah Purdum Reed Riddle Jesper Sollerman Anastasia Wei Mansi M. Kasliwal
This is my paper

Pith reviewed 2026-07-02 17:00 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.HEcs.LG
keywords transient classificationmultimodal machine learningZwicky Transient Facilityreal-time classificationlight curvesastronomical alertsORACLEhierarchical classification
0
0 comments X

The pith

Adding metadata and images to light curves raises real-time transient classification accuracy by up to 40 percent on ZTF data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that models fusing light curves with metadata and images outperform light-curve-only baselines for early classification of astronomical transients and variables. On real ZTF Bright Transient Survey observations the best multimodal model reaches a macro F1 of 0.73 while a light-curve-plus-metadata variant attains 0.88 on simulated ELAsTiCC data. These gains matter because current and future surveys produce hundreds of thousands of nightly alerts that require rapid triage for follow-up resources. The work deploys the models on the live ZTF alert stream and measures performance-throughput trade-offs to show practicality for higher-volume surveys such as LSST.

Core claim

The central claim is that ORACLE-2 hierarchical classifiers that combine light-curve time series, metadata features, and image data deliver consistently higher macro F1 scores than unimodal light-curve models, with the largest improvements occurring at the earliest epochs when light curves remain sparse. On ZTF BTS data the full multimodal ORACLE-2 Omni model achieves 0.73 macro F1, an improvement of up to 11 percent over light-curve-plus-metadata and up to 40 percent over light-curve-only versions. A light-curve-plus-metadata model trained on the ELAsTiCC simulation reaches 0.88 macro F1, matching other state-of-the-art results while remaining deployable on the actual ZTF alert stream.

What carries the argument

The ORACLE-2 models that perform hierarchical classification by fusing light-curve time series with metadata and image inputs to output real-time class probabilities.

Load-bearing premise

The training and test distributions from ZTF BTS and ELAsTiCC simulations are representative enough of future live alert streams that the reported F1 gains will hold under different class priors and alert rates.

What would settle it

A deployment test on new live ZTF alerts in which the multimodal model's macro F1 on a held-out sample of confirmed transients falls below the light-curve-only baseline.

Figures

Figures reproduced from arXiv: 2607.00228 by Adam A. Miller, Anastasia Wei, Antoine Le Calloch, Ashish A. Mahabal, Frank J. Masci, Jesper Sollerman, Joahan Castaneda Jaimes, Josiah Purdum, Mansi M. Kasliwal, Matthew J. Graham, Michael W. Coughlin, Nabeel Rehemtulla, Reed Riddle, Sushant Sharma Chaudhary, Theophile Jegou Du Laz, Ved G. Shah.

Figure 1
Figure 1. Figure 1: Example light curves and metadata from the BTS dataset, along with the g, r, and i band images from PanSTARRS-1 for all seven leaf classes in our taxonomy. plete extragalactic transient survey utilizing ZTF data. BTS ran from April 2018 through the end of 2024 and produced a large (>10,000 sources), highly complete (∼ 95% completeness) spectroscopic sample of extra￾galactic transients with mpeak ≲ 18.5 mag… view at source ↗
Figure 2
Figure 2. Figure 2: Representative images for different classes in the BTS dataset. Each row corresponds to a different object, while columns (left to right) show the PS1 gri image, the ZTF g-band, the ZTF r-band, and the ZTF i-band reference images. 2.1.1. Additional Features and Modalities In addition to the light curves and metadata from the alerts, we augment our dataset with additional modali￾ties and contextual informat… view at source ↗
Figure 3
Figure 3. Figure 3: Taxonomy used for classifying sources from the Bright Transient Survey dataset. ets, thus removing reliance on any external dependen￾cies. Instead of including the science, reference, and dif￾ference cutouts provided individually in each alert, we focus on the reference images and create a 63 × 63 × 3 (height × width × channels) tensor of ZTF gri reference images. The image fed to the model only passes the… view at source ↗
Figure 4
Figure 4. Figure 4: High level overview of the architecture and I/O for the ORACLE-2 family of models. The time series backbone uses a recurrent neural network (RNN), the metadata back￾bone uses a multi-layer perceptron (MLP), and the image backbone uses a convolutional neural network (CNN). 2025; Shah et al. 2025; Li et al. 2025; Townsend et al. 2026, among others). Broadly, these models adopt one of two design paradigms: 1.… view at source ↗
Figure 5
Figure 5. Figure 5: This network forms the time series backbone used for the entire ORACLE-2 family of models. Not only are these models effective for classification given enough training data (see Section 5), but they also act as strong starting points to train models for other sur￾veys via transfer learning (Gupta et al. 2025). 3.2. Oracle-2 In addition to light curves, alert packets from large time-domain surveys contain a… view at source ↗
Figure 6
Figure 6. Figure 6: Depth 2 F1 scores as a function of time for the single channel ZTF and triple channel Pan-STARRS-1 ORACLE-2 Omni models. models and consider the performance-throughput trade￾offs. Section 5.3 discusses common failure modes for our models. 5.1. Classification Performance First, we define the precision and recall for a class as Precision = TP/(TP+FP) and Recall = TP/ (TP+FN), where TP, FP, and FN are the num… view at source ↗
Figure 7
Figure 7. Figure 7: Time evolution of the macro F1 scores at depth 1 (left) and depth 2 (right) for the Image Backbone (image only), ORACLE-2 Lite (light curve only), ORACLE-2 (light curve + metadata), and ORACLE-2 Omni (light curve + metadata + images) models for the ZTF BTS dataset. Depth 1 distinguishes between transient and persistent sources while depth 2 allows for more granular classification between supernova sub-type… view at source ↗
Figure 8
Figure 8. Figure 8: Time evolution of the class F1 scores across all 9 classes in our taxonomy for the Image Backbone (image only), ORACLE-2 Lite (light curve only), ORACLE-2 (light curve + metadata), and ORACLE-2 Omni (light curve + metadata + images) models for the Bright Transient Survey (BTS) dataset. At 1024 days after the first detection, we report a 19-way (leaf depth) macro F1 score of 0.88 ± 0.00 and 0.83 ± 0.00 for … view at source ↗
Figure 9
Figure 9. Figure 9: Depth 2 confusion matrices for the ORACLE-2 Omni model, 1024 days after the first detection, normal￾ized by the predicted/photometric class (top) and the true/spectroscopic class (bottom). 2.10 GHz with 188 GB of memory, and an NVIDIA A100 graphics processing unit (GPU) with 40 GB of video memory18. The inference throughput for each model using both the CPU and GPU, along with its parame￾ter count, is repo… view at source ↗
Figure 10
Figure 10. Figure 10: Depth 1 (top) and depth 2 (bottom) confusion matrices for the ORACLE-2 Lite (left; uses light curves only), ORACLE-2 (middle; uses light curves + metadata), and ORACLE-2 Omni (right; uses light curves + metadata + images) models, normalized by the true/spectroscopic class, 1 day after the first detection. 5.3. Failure mode analysis for BTS As we have stated in Section 2.1, we do not apply any quality cuts… view at source ↗
Figure 11
Figure 11. Figure 11: Depth 1 (top) and depth 2 (bottom) confusion matrices for the ORACLE-2 Lite (left; uses light curves only), ORACLE-2 (middle; uses light curves + metadata), and ORACLE-2 Omni (right; uses light curves + metadata + images) models, normalized by the true/spectroscopic class, 8 days after the first detection. 1 2 4 8 16 32 64 128 256 512 1024 Days since first detection 0.88 0.90 0.92 0.94 0.96 0.98 1.00 Macr… view at source ↗
Figure 12
Figure 12. Figure 12: Time evolution of the macro F1 scores at depth 1 (left), depth 2 (middle), and leaf depth (right) for the ORACLE-2 Lite (light curve only) and ORACLE-2 (light curve + metadata) models for the ELAsTiCC dataset. (similar to the one detailed in Perley et al. 2020) pro￾vided by BOOM, consuming alerts via Kafka and upload￾ing classifications to Fritz in near real-time. Inference is re-run on each new alert for… view at source ↗
Figure 13
Figure 13. Figure 13: Top: Confusion matrix for the ORACLE-2 model trained on the ELAsTiCC dataset, 1024 days after the first detection. Bottom: Difference in the confusion matrices of the ORACLE-2 and ORACLE-1 models, 1024 days after the first detection. ORACLE-2 consistently shows improvements in power (blue) along the diagonal elements and reduction in power (red) for off-diagonal elements, highlighting better agreement bet… view at source ↗
Figure 14
Figure 14. Figure 14: Asymmetric classification matrix, normalized by the true class, illustrating the classifications for sources which do not neatly fit into our taxonomy for BTS. During this period, ORACLE-2 Omni classified 344 new sources19 which also had a spectroscopic label, achiev￾ing a depth 1 time-averaged macro F1-score of 0.88 and accuracy of 0.99. At depth 2, we report a time-averaged macro F1-score of 0.55 and ac… view at source ↗
Figure 15
Figure 15. Figure 15: Time evolution of the class F1 scores for the ORACLE-2 Lite (light curve only) and ORACLE-2 (light curve + metadata) models for all 19 leaf classes in the ELAsTiCC dataset [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗
read the original abstract

Modern time-domain surveys such as the Zwicky Transient Facility (ZTF) generate hundreds of thousands of alerts each night, making real-time decisions for follow-up observations a central challenge in time-domain astronomy. Robust early classification is crucial for making informed decisions, but is hindered by sparse light curves and degeneracies between classes. In this work, we leverage multimodality to substantially improve real-time classification and demonstrate the practicality of our approach by deploying our model on the ZTF alert stream. Building on the Online Ranked Astrophysical CLass Estimator (ORACLE), we introduce the ORACLE-2 models, which combine light curves, metadata, and images for real-time hierarchical classification. Using both real and simulated datasets, we show that incorporating additional modalities consistently improves classification performance. On observations from ZTF's Bright Transient Survey, our best-performing model, ORACLE-2 Omni, achieves a macro F1 score of 0.73 -- an improvement of up to 11% over models using light curves and metadata alone, and up to 40% over light-curve-only models, with the strongest gains realized at early times. To demonstrate applicability to the Legacy Survey of Space and Time, which will increase alert volume by more than an order of magnitude, we train a light curve + metadata variant on the simulated ELAsTiCC dataset. This model achieves a macro F1 score of 0.88, an improvement of up to 13% over the light-curve-only variant, matching the performance of other state-of-the-art models. Finally, we quantify the trade-offs between performance and throughput, identifying regimes where multimodal approaches offer the greatest benefit. These results show that combining multiple modalities improves early-time classification, enabling more effective triage of high-volume alert streams for current and future time-domain surveys.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces ORACLE-2, an extension of the ORACLE framework, for real-time hierarchical classification of transients and variables using multimodal inputs (light curves, metadata, and images) from ZTF alerts. It reports consistent macro F1 gains on real ZTF BTS data (0.73 for the Omni model, up to 40% over light-curve-only) and on simulated ELAsTiCC data (0.88 for the light-curve+metadata variant), with largest improvements at early times, quantifies performance-throughput trade-offs, and states that the model has been deployed on the live ZTF alert stream.

Significance. If the multimodal gains hold under operational conditions, the approach would offer a concrete, deployable method for improving early triage in high-volume alert streams, directly relevant to ZTF operations and scalable to LSST; the empirical results on both real and simulated data plus the throughput analysis are strengths that would support adoption if distribution-shift concerns are addressed.

major comments (2)
  1. [Abstract] Abstract and deployment description: the central claim that the reported F1 gains (0.73 on BTS, 0.88 on ELAsTiCC) will translate to live ZTF operations rests on the untested assumption that BTS/ELAsTiCC class priors, alert cadence, magnitude limits, and host-galaxy properties match the operational stream; no quantitative comparison or re-evaluation on a held-out live segment is supplied.
  2. [Abstract] Abstract: the statement that ORACLE-2 Omni achieves up to 40% improvement over light-curve-only models is presented without accompanying error bars, class-wise breakdown, or analysis of whether post-hoc class-imbalance handling or validation-split choices affect the macro F1 numbers.
minor comments (1)
  1. The methods section should explicitly state the exact input preprocessing steps for images and metadata and the hierarchical decision thresholds used at inference time.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We address each of the major comments below and propose revisions where appropriate to strengthen the presentation of our results.

read point-by-point responses

  1. Referee: [Abstract] Abstract and deployment description: the central claim that the reported F1 gains (0.73 on BTS, 0.88 on ELAsTiCC) will translate to live ZTF operations rests on the untested assumption that BTS/ELAsTiCC class priors, alert cadence, magnitude limits, and host-galaxy properties match the operational stream; no quantitative comparison or re-evaluation on a held-out live segment is supplied.

    Authors: We acknowledge that the BTS dataset, while consisting of real ZTF observations, represents a specific bright transient sample and may not fully capture the diversity of the entire live ZTF alert stream. Similarly, ELAsTiCC is a simulation. Our live deployment demonstrates that the model can run in real time on the ZTF stream, but we agree that without a held-out live evaluation with labels, direct translation of the F1 scores cannot be quantitatively verified. We will revise the abstract to clarify that the performance metrics are reported on BTS and ELAsTiCC, and that the deployment serves to show operational feasibility rather than to claim identical performance on the full stream. We will also add a section discussing potential distribution shifts. revision: yes

  2. Referee: [Abstract] Abstract: the statement that ORACLE-2 Omni achieves up to 40% improvement over light-curve-only models is presented without accompanying error bars, class-wise breakdown, or analysis of whether post-hoc class-imbalance handling or validation-split choices affect the macro F1 numbers.

    Authors: The 'up to 40%' figure refers to the maximum relative improvement observed across different time bins or configurations on the BTS dataset. To address this, we will include error bars on the macro F1 scores in the abstract and figures, expand the results section with a class-wise performance breakdown, and add an appendix or subsection analyzing the sensitivity to class-imbalance handling methods and different validation splits to confirm the robustness of the reported improvements. revision: yes

Circularity Check

0 steps flagged

No circularity: all claims are empirical measurements on held-out data

full rationale

The paper reports macro F1 scores from training and evaluating hierarchical classifiers on fixed train/test splits of ZTF BTS and ELAsTiCC data. No equations, first-principles derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the chain. The ORACLE base model is cited as prior work, but the multimodal gains (0.73 and 0.88 F1) are direct empirical outcomes on independent test sets, not reductions to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard supervised-learning assumptions plus the representativeness of the two evaluation datasets; no new physical entities or ad-hoc constants are introduced.

axioms (2)
  • domain assumption Training and test data are drawn from distributions representative of future operational alert streams
    Required for the reported F1 gains to translate to live deployment; invoked implicitly when claiming applicability to LSST.
  • standard math Standard neural-network training assumptions (i.i.d. samples, fixed class taxonomy)
    Background assumption for any supervised classifier; not stated explicitly in abstract.

pith-pipeline@v0.9.1-grok · 5956 in / 1418 out tokens · 21766 ms · 2026-07-02T17:00:20.925590+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

103 extracted references · 93 canonical work pages · 12 internal anchors

  1. [1]

    The Wide Field Infrared Survey Telescope: 100 Hubbles for the 2020s

    Akeson, R., Armus, L., Bachelet, E., et al. 2019, arXiv e-prints, arXiv:1902.05569, doi: 10.48550/arXiv.1902.05569

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1902.05569 2019

  2. [2]

    J., Stern, D., Noirot, G., et al

    Assef, R. J., Stern, D., Noirot, G., et al. 2018, The Astrophysical Journal Supplemental Series, 234, 23, doi: 10.3847/1538-4365/aaa00a Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068 Astropy Collaboration, Price-Whelan, A. M., Sip˝ ocz, B. M., et al. 2018, ApJ, 156, 123, doi: 10.3847...

    work page doi:10.3847/1538-4365/aaa00a 2018

  3. [3]

    Neural Machine Translation by Jointly Learning to Align and Translate

    Bahdanau, D., Cho, K., & Bengio, Y. 2014, arXiv e-prints, arXiv:1409.0473, doi: 10.48550/arXiv.1409.0473

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1409.0473 2014

  4. [4]

    C., Kulkarni, S

    Bellm, E. C., Kulkarni, S. R., Graham, M. J., et al. 2019, PASP, 131, 018002, doi: 10.1088/1538-3873/aaecbe

    work page doi:10.1088/1538-3873/aaecbe 2019

  5. [5]

    Lord, N. A. 2019, arXiv e-prints, arXiv:1912.09393, doi: 10.48550/arXiv.1912.09393

    work page doi:10.48550/arxiv.1912.09393 2019

  6. [6]

    2020, Experiment Tracking with Weights and Biases

    Biewald, L. 2020, Experiment Tracking with Weights and Biases. https://www.wandb.com/

    2020

  7. [7]

    D., Walters, R., et al

    Blagorodnova, N., Neill, J. D., Walters, R., et al. 2018, PASP, 130, 035003, doi: 10.1088/1538-3873/aaa53f

    work page doi:10.1088/1538-3873/aaa53f 2018

  8. [8]

    2019, AJ, 158, 257, doi: 10.3847/1538-3881/ab5182 —

    Boone, K. 2019, AJ, 158, 257, doi: 10.3847/1538-3881/ab5182 —. 2021, AJ, 162, 275, doi: 10.3847/1538-3881/ac2a2d

    work page doi:10.3847/1538-3881/ab5182 2019

  9. [9]

    2024, A&A, 689, A289, doi: 10.1051/0004-6361/202449475 C´ adiz-Leyton, M., Cabrera-Vives, G., Protopapas, P.,

    Cabrera-Vives, G., Moreno-Cartagena, D., Astorga, N., et al. 2024, A&A, 689, A289, doi: 10.1051/0004-6361/202449475 C´ adiz-Leyton, M., Cabrera-Vives, G., Protopapas, P.,

    work page doi:10.1051/0004-6361/202449475 2024

  10. [10]

    2025, arXiv e-prints, arXiv:2507.12611, doi: 10.48550/arXiv.2507.12611

    Moreno-Cartagena, D., & Becker, I. 2025, arXiv e-prints, arXiv:2507.12611, doi: 10.48550/arXiv.2507.12611

    work page doi:10.48550/arxiv.2507.12611 2025

  11. [11]

    Chaini, S., & Kumar, S. S. 2020, arXiv e-prints, arXiv:2006.12333, doi: 10.48550/arXiv.2006.12333

    work page doi:10.48550/arxiv.2006.12333 2020

  12. [12]

    The Pan-STARRS1 Surveys

    Chambers, K. C., Magnier, E. A., Metcalfe, N., et al. 2016, arXiv e-prints, arXiv:1612.05560, doi: 10.48550/arXiv.1612.05560

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1612.05560 2016

  13. [13]

    2020, The Astrophysical Journal Supplement Series, 249, 18, doi: 10.3847/1538-4365/ab9cae

    Chen, X., Wang, S., Deng, L., et al. 2020, The Astrophysical Journal Supplement Series, 249, 18, doi: 10.3847/1538-4365/ab9cae

    work page doi:10.3847/1538-4365/ab9cae 2020

  14. [14]

    H., Yan, L., Kangas, T., et al

    Chen, Z. H., Yan, L., Kangas, T., et al. 2023, ApJ, 943, 41, doi: 10.3847/1538-4357/aca161

    work page doi:10.3847/1538-4357/aca161 2023

  15. [15]

    On the Properties of Neural Machine Translation: Encoder-Decoder Approaches

    Cho, K., van Merrienboer, B., Bahdanau, D., & Bengio, Y. 2014, arXiv e-prints, arXiv:1409.1259, doi: 10.48550/arXiv.1409.1259

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1409.1259 2014

  16. [16]

    W., Bloom, J

    Coughlin, M. W., Bloom, J. S., Nir, G., et al. 2023, ApJS, 267, 31, doi: 10.3847/1538-4365/acdee1

    work page doi:10.3847/1538-4365/acdee1 2023

  17. [17]

    M., Wright, E

    Cutri, R. M., Wright, E. L., Conrow, T., et al. 2021, VizieR Online Data Catalog: AllWISE Data Release (Cutri+ 2013), VizieR On-line Data Catalog: II/328. Originally published in: IPAC/Caltech (2013) de Soto, K. M., Villar, V. A., Berger, E., et al. 2024, ApJ, 974, 169, doi: 10.3847/1538-4357/ad6a4f

    work page doi:10.3847/1538-4357/ad6a4f 2021

  18. [18]

    M., Riddle, R., et al

    Dekany, R., Smith, R. M., Riddle, R., et al. 2020, PASP, 132, 038001, doi: 10.1088/1538-3873/ab4ca2 Della Valle, M., & Izzo, L. 2020, A&A Rv, 28, 3, doi: 10.1007/s00159-020-0124-6

    work page doi:10.1088/1538-3873/ab4ca2 2020

  19. [19]

    2009, in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–255, doi: 10.1109/CVPR.2009.5206848 DES Collaboration, Abbott, T

    Deng, J., Dong, W., Socher, R., et al. 2009, in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–255, doi: 10.1109/CVPR.2009.5206848 DES Collaboration, Abbott, T. M. C., Acevedo, M., et al. 2025, The Dark Energy Survey: Cosmology Results With 1500 New High-redshift Type Ia Supernovae Using The Full 5-year Dataset. https://arxiv.org/abs...

    work page doi:10.1109/cvpr.2009.5206848 2009

  20. [20]

    A., Mahabal, A., Masci, F

    Duev, D. A., Mahabal, A., Masci, F. J., et al. 2019, MNRAS, 489, 3582, doi: 10.1093/mnras/stz2357

    work page doi:10.1093/mnras/stz2357 2019

  21. [21]

    J., & Mandel, K

    Foley, R. J., & Mandel, K. 2013, ApJ, 778, 167, doi: 10.1088/0004-637X/778/2/167

    work page doi:10.1088/0004-637x/778/2/167 2013

  22. [22]

    A., Sharma, Y., et al

    Fremling, C., Miller, A. A., Sharma, Y., et al. 2020, ApJ, 895, 32, doi: 10.3847/1538-4357/ab8943

    work page doi:10.3847/1538-4357/ab8943 2020

  23. [23]

    J., Coughlin, M

    Fremling, C., Hall, X. J., Coughlin, M. W., et al. 2021, ApJL, 917, L2, doi: 10.3847/2041-8213/ac116f

    work page doi:10.3847/2041-8213/ac116f 2021

  24. [24]

    2025, ApJ, 992, 158, doi: 10.3847/1538-4357/adff4e

    Frohmaier, C., Vincenzi, M., Sullivan, M., et al. 2025, ApJ, 992, 158, doi: 10.3847/1538-4357/adff4e

    work page doi:10.3847/1538-4357/adff4e 2025

  25. [25]

    I., & Aleo, P

    Gagliano, A., Contardo, G., Foreman-Mackey, D., Malz, A. I., & Aleo, P. D. 2023, ApJ, 954, 6, doi: 10.3847/1538-4357/ace326

    work page doi:10.3847/1538-4357/ace326 2023

  26. [26]

    2021, ApJ, 908, 170, doi: 10.3847/1538-4357/abd02b

    Gagliano, A., Narayan, G., Engel, A., Carrasco Kind, M., & LSST Dark Energy Science Collaboration. 2021, ApJ, 908, 170, doi: 10.3847/1538-4357/abd02b

    work page doi:10.3847/1538-4357/abd02b 2021

  27. [27]

    T., Shen, Y., & Villar, V

    Gagliano, A. T., Shen, Y., & Villar, V. A. 2025, arXiv e-prints, arXiv:2512.04145, doi: 10.48550/arXiv.2512.04145

    work page doi:10.48550/arxiv.2512.04145 2025

  28. [28]

    2017, in Handbook of Supernovae, ed

    Gal-Yam, A. 2017, in Handbook of Supernovae, ed. A. W. Alsabti & P. Murdin, 195, doi: 10.1007/978-3-319-21846-5 35

    work page doi:10.1007/978-3-319-21846-5 2017

  29. [29]

    2019, Annual Review of Astronomy and Astrophysics, 57, 305–333, doi: 10.1146/annurev-astro-081817-051819

    Gal-Yam, A. 2019, Annual Review of Astronomy and Astrophysics, 57, 305–333, doi: 10.1146/annurev-astro-081817-051819

    work page doi:10.1146/annurev-astro-081817-051819 2019

  30. [30]

    K., et al

    Gomez, S., Berger, E., Blanchard, P. K., et al. 2020, ApJ, 904, 74, doi: 10.3847/1538-4357/abbf49

    work page doi:10.3847/1538-4357/abbf49 2020

  31. [31]

    2024, MNRAS, 535, 471, doi: 10.1093/mnras/stae2270

    Gomez, S., Nicholl, M., Berger, E., et al. 2024, MNRAS, 535, 471, doi: 10.1093/mnras/stae2270

    work page doi:10.1093/mnras/stae2270 2024

  32. [32]

    J., Kulkarni, S

    Graham, M. J., Kulkarni, S. R., Bellm, E. C., et al. 2019, PASP, 131, 078001, doi: 10.1088/1538-3873/ab006c Multimodality for real-time classification of transients23

    work page doi:10.1088/1538-3873/ab006c 2019

  33. [33]

    2025, MNRAS, 542, L132, doi: 10.1093/mnrasl/slaf074

    Gupta, R., Muthukrishna, D., Rehemtulla, N., & Shah, V. 2025, MNRAS, 542, L132, doi: 10.1093/mnrasl/slaf074

    work page doi:10.1093/mnrasl/slaf074 2025

  34. [34]

    2008, Exploring network structure, dynamics, and function using

    Hagberg, A., Swart, P., & S Chult, D. 2008, Exploring network structure, dynamics, and function using

    2008

  35. [35]

    A., Nazaryan, T

    Hakobyan, A. A., Nazaryan, T. A., Adibekyan, V. Z., et al. 2014, MNRAS, 444, 2428, doi: 10.1093/mnras/stu1598

    work page doi:10.1093/mnras/stu1598 2014

  36. [36]

    R., Millman, K

    Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2

    work page doi:10.1038/s41586-020-2649-2 2020

  37. [37]

    2016, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1, doi: 10.1109/CVPR.2016.90

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1, doi: 10.1109/CVPR.2016.90

    work page doi:10.1109/cvpr.2016.90 2016

  38. [38]

    Gaussian Error Linear Units (GELUs)

    Hendrycks, D., & Gimpel, K. 2016, arXiv e-prints, arXiv:1606.08415, doi: 10.48550/arXiv.1606.08415

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1606.08415 2016

  39. [39]

    A., Sollerman, J., et al

    Hinds, K.-R., Perley, D. A., Sollerman, J., et al. 2025, MNRAS, 541, 135, doi: 10.1093/mnras/staf888

    work page doi:10.1093/mnras/staf888 2025

  40. [40]

    Distilling the Knowledge in a Neural Network

    Hinton, G. E., Vinyals, O., & Dean, J. 2015, ArXiv, abs/1503.02531. https://api.semanticscholar.org/CorpusID:7200347

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  41. [41]

    Howell, D. A. 2011, Nature Communications, 2, doi: 10.1038/ncomms1344

    work page doi:10.1038/ncomms1344 2011

  42. [42]

    Hunter, J. D. 2007, Computing In Science & Engineering, 9, 90 Inc., P. T. 2015, Collaborative data science, Montreal, QC: Plotly Technologies Inc. https://plot.ly Ivezi´ c,ˇZ., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111, doi: 10.3847/1538-4357/ab042c Jegou du Laz, T., Coughlin, M. W., Bachant, P., et al. 2025, arXiv e-prints, arXiv:2511.00164, ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.3847/1538-4357/ab042c 2007

  43. [43]

    2025, arXiv e-prints, arXiv:2507.16088, doi: 10.48550/arXiv.2507.16088

    Junell, A., Sasli, A., Fontinele Nunes, F., et al. 2025, arXiv e-prints, arXiv:2507.16088, doi: 10.48550/arXiv.2507.16088

    work page doi:10.48550/arxiv.2507.16088 2025

  44. [44]

    E., et al

    Kaiser, N., Aussel, H., Burke, B. E., et al. 2002, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4836, Survey and Other Telescope Technologies and Discoveries, ed. J. A. Tyson & S. Wolff, 154–164, doi: 10.1117/12.457365

    work page doi:10.1117/12.457365 2002

  45. [45]

    M., Bellm, E., & Graham, M

    Kasliwal, M. M., Bellm, E., & Graham, M. 2025, Transient Name Server AstroNote, 238, 1

    2025

  46. [46]

    L., & Kirshner, R

    Kelly, P. L., & Kirshner, R. P. 2012, ApJ, 759, 107, doi: 10.1088/0004-637X/759/2/107

    work page doi:10.1088/0004-637x/759/2/107 2012

  47. [47]

    D., et al

    Kim, Y.-L., Rigault, M., Neill, J. D., et al. 2022, PASP, 134, 024505, doi: 10.1088/1538-3873/ac50a0

    work page doi:10.1088/1538-3873/ac50a0 2022

  48. [48]

    Adam: A Method for Stochastic Optimization

    Kingma, D. P., & Ba, J. 2014, CoRR, abs/1412.6980. https://api.semanticscholar.org/CorpusID:6628106

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  49. [49]

    2023, in American Astronomical Society Meeting Abstracts, Vol

    Knop, R., & ELAsTiCC Team. 2023, in American Astronomical Society Meeting Abstracts, Vol. 241, American Astronomical Society Meeting Abstracts #241, 117.02

    2023

  50. [50]

    Kulkarni, S. R. 2020, arXiv e-prints, arXiv:2004.03511. https://arxiv.org/abs/2004.03511

    work page arXiv 2020

  51. [51]

    M., Corbett, H., Galliher, N

    Law, N. M., Corbett, H., Galliher, N. W., et al. 2022, PASP, 134, 035003, doi: 10.1088/1538-3873/ac4811

    work page doi:10.1088/1538-3873/ac4811 2022

  52. [52]

    1989, in Advances in Neural Information Processing Systems, ed

    LeCun, Y., Boser, B., Denker, J., et al. 1989, in Advances in Neural Information Processing Systems, ed. D. Touretzky, Vol. 2 (Morgan-Kaufmann). https://proceedings.neurips.cc/paper files/paper/1989/ file/53c3bce66e43be4f209556518c2fcb54-Paper.pdf

    1989

  53. [53]

    2025, arXiv e-prints, arXiv:2510.06200, doi: 10.48550/arXiv.2510.06200

    Li, W., Chen, H.-Y., Rehemtulla, N., et al. 2025, arXiv e-prints, arXiv:2510.06200, doi: 10.48550/arXiv.2510.06200

    work page doi:10.48550/arxiv.2510.06200 2025

  54. [54]

    , keywords =

    Lintott, C., Schawinski, K., Bamford, S., et al. 2011, MNRAS, 410, 166, doi: 10.1111/j.1365-2966.2010.17432.x

    work page doi:10.1111/j.1365-2966.2010.17432.x 2011

  55. [55]

    2008 , month =

    Lintott, C. J., Schawinski, K., Slosar, A., et al. 2008, MNRAS, 389, 1179, doi: 10.1111/j.1365-2966.2008.13689.x

    work page doi:10.1111/j.1365-2966.2008.13689.x 2008

  56. [56]

    2022, arXiv e-prints, arXiv:2201.03545, doi: 10.48550/arXiv.2201.03545 LSST Dark Energy Science Collaboration, Aubourg, E.,

    Liu, Z., Mao, H., Wu, C.-Y., et al. 2022, arXiv e-prints, arXiv:2201.03545, doi: 10.48550/arXiv.2201.03545 LSST Dark Energy Science Collaboration, Aubourg, E.,

    work page doi:10.48550/arxiv.2201.03545 2022

  57. [57]

    2026, arXiv e-prints, arXiv:2601.14235, doi: 10.48550/arXiv.2601.14235

    Avestruz, C., et al. 2026, arXiv e-prints, arXiv:2601.14235, doi: 10.48550/arXiv.2601.14235

    work page doi:10.48550/arxiv.2601.14235 2026

  58. [58]

    2015, ApJ, 804, 90, doi: 10.1088/0004-637X/804/2/90

    Lunnan, R., Chornock, R., Berger, E., et al. 2015, ApJ, 804, 90, doi: 10.1088/0004-637X/804/2/90

    work page doi:10.1088/0004-637x/804/2/90 2015

  59. [59]

    2023, in American Astronomical Society Meeting Abstracts, Vol

    Malanchev, K. 2023, in American Astronomical Society Meeting Abstracts, Vol. 241, American Astronomical Society Meeting Abstracts #241, 117.03

    2023

  60. [60]

    I., & ELAsTiCC Team

    Malz, A. I., & ELAsTiCC Team. 2023, in American Astronomical Society Meeting Abstracts, Vol. 241, American Astronomical Society Meeting Abstracts #241, 117.04

    2023

  61. [61]

    J., Laher, R

    Masci, F. J., Laher, R. R., Rusholme, B., et al. 2019, PASP, 131, 018003, doi: 10.1088/1538-3873/aae8ac

    work page internal anchor Pith review doi:10.1088/1538-3873/aae8ac 2019

  62. [62]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    McInnes, L., Healy, J., & Melville, J. 2018, arXiv e-prints, arXiv:1802.03426, doi: 10.48550/arXiv.1802.03426

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1802.03426 2018

  63. [63]

    A., Abrams, N

    Miller, A. A., Abrams, N. S., Aldering, G., et al. 2025, PASP, 137, 094204, doi: 10.1088/1538-3873/ae02c5 M¨ oller, A., & de Boissi` ere, T. 2020, MNRAS, 491, 4277, doi: 10.1093/mnras/stz3312

    work page doi:10.1088/1538-3873/ae02c5 2025

  64. [64]

    2025, A&A, 703, A41, doi: 10.1051/0004-6361/202554289

    Moreno-Cartagena, D., Protopapas, P., Cabrera-Vives, G., et al. 2025, A&A, 703, A41, doi: 10.1051/0004-6361/202554289

    work page doi:10.1051/0004-6361/202554289 2025

  65. [65]

    S., Biswas, R., & Hloˇ zek, R

    Muthukrishna, D., Narayan, G., Mandel, K. S., Biswas, R., & Hloˇ zek, R. 2019, PASP, 131, 118002, doi: 10.1088/1538-3873/ab1609 24Shah et al

    work page doi:10.1088/1538-3873/ab1609 2019

  66. [66]

    2023, in American Astronomical Society Meeting Abstracts, Vol

    Narayan, G., & ELAsTiCC Team. 2023, in American Astronomical Society Meeting Abstracts, Vol. 241, American Astronomical Society Meeting Abstracts #241, 117.01

    2023

  67. [67]

    D., Sullivan, M., Gal-Yam, A., et al

    Neill, J. D., Sullivan, M., Gal-Yam, A., et al. 2011, ApJ, 727, 15, doi: 10.1088/0004-637X/727/1/15

    work page doi:10.1088/0004-637x/727/1/15 2011

  68. [68]

    M., Assef, R

    Padovani, P., Alexander, D. M., Assef, R. J., et al. 2017, The Astronomy and Astrophysics Review, 25, doi: 10.1007/s00159-017-0102-9

    work page doi:10.1007/s00159-017-0102-9 2017

  69. [69]

    AION-1: Omnimodal Foundation Model for Astronomical Sciences

    Parker, L., Lanusse, F., Shen, J., et al. 2025, arXiv e-prints, arXiv:2510.17960, doi: 10.48550/arXiv.2510.17960

    work page doi:10.48550/arxiv.2510.17960 2025

  70. [70]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library

    Paszke, A., Gross, S., Massa, F., et al. 2019, arXiv e-prints, arXiv:1912.01703, doi: 10.48550/arXiv.1912.01703

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1912.01703 2019

  71. [71]

    2011, Journal of Machine Learning Research, 12, 2825

    Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, Journal of Machine Learning Research, 12, 2825

    2011

  72. [72]

    A., Quimby, R

    Perley, D. A., Quimby, R. M., Yan, L., et al. 2016, ApJ, 830, 13, doi: 10.3847/0004-637X/830/1/13

    work page doi:10.3847/0004-637x/830/1/13 2016

  73. [73]

    A., Fremling, C., Sollerman, J., et al

    Perley, D. A., Fremling, C., Sollerman, J., et al. 2020, ApJ, 904, 35, doi: 10.3847/1538-4357/abbd98

    work page doi:10.3847/1538-4357/abbd98 2020

  74. [74]

    J., Ashall, C., James, P

    Prentice, S. J., Ashall, C., James, P. A., et al. 2019, MNRAS, 485, 1559, doi: 10.1093/mnras/sty3399

    work page doi:10.1093/mnras/sty3399 2019

  75. [75]

    2021, AJ, 162, 67, doi: 10.3847/1538-3881/ac0824

    Qu, H., Sako, M., M¨ oller, A., & Doux, C. 2021, AJ, 162, 67, doi: 10.3847/1538-3881/ac0824

    work page doi:10.3847/1538-3881/ac0824 2021

  76. [76]

    W., Miller, A

    Rehemtulla, N., Coughlin, M. W., Miller, A. A., & du Laz, T. J. 2025a, Nature Astronomy, 9, 1764, doi: 10.1038/s41550-025-02720-6

    work page doi:10.1038/s41550-025-02720-6

  77. [77]

    A., Jegou Du Laz, T., et al

    Rehemtulla, N., Miller, A. A., Jegou Du Laz, T., et al. 2024, ApJ, 972, 7, doi: 10.3847/1538-4357/ad5666

    work page doi:10.3847/1538-4357/ad5666 2024

  78. [78]

    V., Singh, A., et al

    Rehemtulla, N., Jacobson-Gal´ an, W. V., Singh, A., et al. 2025b, ApJ, 985, 241, doi: 10.3847/1538-4357/adcf1e

    work page doi:10.3847/1538-4357/adcf1e

  79. [79]

    A., Walmsley, M., et al

    Rehemtulla, N., Miller, A. A., Walmsley, M., et al. 2026, PASP, 138, 034503, doi: 10.1088/1538-3873/ae50bc

    work page doi:10.1088/1538-3873/ae50bc 2026

  80. [80]

    D., Blagorodnova, N., et al

    Rigault, M., Neill, J. D., Blagorodnova, N., et al. 2019, A&A, 627, A115, doi: 10.1051/0004-6361/201935344

    work page doi:10.1051/0004-6361/201935344 2019

Showing first 80 references.