SELDON: Supernova Explosions Learned by Deep ODE Networks
| Authors | Jiezhong Wu et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2603.04392 |
| Download | |
| Categories | cs.LG |
Abstract
The discovery rate of optical transients will explode to 10 million public alerts per night once the Vera C. Rubin Observatory's Legacy Survey of Space and Time comes online, overwhelming the traditional physics-based inference pipelines. A continuous-time forecasting AI model is of interest because it can deliver millisecond-scale inference for thousands of objects per day, whereas legacy MCMC codes need hours per object. In this paper, we propose SELDON, a new continuous-time variational autoencoder for panels of sparse and irregularly time-sampled (gappy) astrophysical light curves that are nonstationary, heteroscedastic, and inherently dependent. SELDON combines a masked GRU-ODE encoder with a latent neural ODE propagator and an interpretable Gaussian-basis decoder. The encoder learns to summarize panels of imbalanced and correlated data even when only a handful of points are observed. The neural ODE then integrates this hidden state forward in continuous time, extrapolating to future unseen epochs. This extrapolated time series is further encoded by deep sets to a latent distribution that is decoded to a weighted sum of Gaussian basis functions, the parameters of which are physically meaningful. Such parameters (e.g., rise time, decay rate, peak flux) directly drive downstream prioritization of spectroscopic follow-up for astrophysical surveys. Beyond astronomy, the architecture of SELDON offers a generic recipe for interpretable and continuous-time sequence modeling in any time domain where data are multivariate, sparse, heteroscedastic, and irregularly spaced.
Engineering Breakdown
Plain English
SELDON is a continuous-time deep learning model designed to forecast astrophysical transient events (like supernovae) from sparse, irregularly-sampled light curve data. The Vera Rubin Observatory will generate 10 million alerts per night starting soon, which would overwhelm traditional physics-based inference pipelines that require hours per object using MCMC methods. The authors propose a neural ODE-based variational autoencoder that can deliver predictions in milliseconds for thousands of objects daily, making it feasible to process the incoming data deluge. The model combines a masked GRU-ODE encoder, a latent neural ODE propagator, and an interpretable Gaussian-basis decoder to handle nonstationary, heteroscedastic, and dependent light curves.
Core Technical Contribution
The key innovation is combining neural ordinary differential equations (ODEs) with a masked GRU encoder specifically designed for irregularly-sampled time series in a variational autoencoder framework. Unlike standard RNNs or Transformers that assume regular time sampling, this approach natively handles gappy astrophysical data where observations arrive at irregular intervals with varying measurement quality. The masked GRU-ODE encoder learns meaningful representations from sparse inputs, while the latent neural ODE propagator can evolve the system state continuously through time without discrete time steps. The Gaussian-basis decoder provides interpretability by decomposing predictions into meaningful astrophysical components, bridging the gap between black-box deep learning and domain-driven physics models.
How It Works
The input is a sparse, irregularly time-sampled light curve—essentially a set of timestamped flux measurements with potentially large gaps and heteroscedastic noise. The masked GRU-ODE encoder processes these observations sequentially, using masking to handle missing timepoints and learning a compact latent representation that captures the transient's evolving state. This latent representation is then evolved forward (or backward) through time using a neural ODE, which learns the underlying continuous-time dynamics without requiring discrete recurrence steps. The decoder, parameterized as a Gaussian basis expansion, reconstructs flux predictions at arbitrary times and provides explainability by showing which basis components (physical processes) explain the light curve behavior. During training, the model optimizes a variational objective that balances reconstruction accuracy against a KL regularizer on the latent space, enabling both accurate forecasting and meaningful latent representations.
Production Impact
For engineers processing real-time astronomical data, SELDON replaces multi-hour MCMC inference pipelines with millisecond-scale predictions, enabling real-time decision-making on which transients to follow up with expensive telescopes. The ability to handle sparse, irregularly-sampled data directly (without interpolation or binning) reduces preprocessing overhead and preserves information loss that occurs in traditional data alignment. However, integration requires careful handling of heteroscedastic noise estimates—the model needs uncertainty quantification to work alongside human astronomers and robotic follow-up systems. Compute costs are significantly lower than Bayesian inference but higher than simple classifiers, typically requiring GPU acceleration; latency scales linearly with batch size rather than the superlinear scaling of MCMC. The interpretable Gaussian-basis decoder is a major advantage for domain adoption, as astronomers can validate that learned components correspond to known physical processes (e.g., shock cooling, radioactive decay).
Limitations and When Not to Use This
The paper abstracts away full details of training procedures, hyperparameter selection, and generalization across diverse transient types, leaving questions about how well the model transfers beyond its training distribution. Neural ODEs can suffer from numerical instability during long time extrapolation or when latent dynamics are highly stiff; the paper doesn't discuss how SELDON handles extreme transient durations or edge cases. The Gaussian-basis decoder, while interpretable, may constrain expressiveness compared to unconstrained neural decoders—there's an implicit assumption that transient light curves decompose into a finite set of smooth basis functions. The model is trained on simulated or archival data; real-time performance under domain shift (instrument changes, observing strategy changes, new transient types) is not validated. Additionally, the paper doesn't address how to handle multi-band photometry (different wavelengths) or how to incorporate auxiliary information like redshift or host galaxy properties.
Research Context
This work builds directly on the neural ODE framework (Chen et al., 2018) and extends it to handle realistic irregularly-sampled time series, following prior work on latent ODEs for continuous-time modeling. It addresses a practical crisis in time-domain astronomy: the discovery rate explosion from next-generation surveys (Rubin, ZTF, ATLAS) has made traditional inference intractable, motivating deep learning approaches for rapid triage. The paper contributes to a growing intersection of physics-informed neural networks and astrophysical data processing, alongside concurrent work on neural network classifiers for transient type identification. Future directions likely include: multi-modal fusion (combining photometry with spectroscopy), hierarchical models that transfer learning across similar transient classes, and uncertainty quantification methods suitable for astronomical decision-making.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::