# Toward an understanding of the properties of neural network approaches for supernovae light curve approximation

**Authors:** Mariia Demianenko, Konstantin Malanchev, Ekaterina Samorodova, Mikhail Sysak, Aleksandr Shiriaev, Denis Derkach, Mikhail Hushchyn  
**arXiv:** [2209.07542](https://arxiv.org/abs/2209.07542)  
**Date:** 2022-09-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.07542v1/

## Key points

- Modern time-domain photometric surveys have revolutionized the field of astronomy by collecting vast amounts of data on various astronomical objects.
- Spectroscopic follow-up is crucial for transients like supernovae, but most objects have never received it.
- Light curves are time series actively used for photometric classification and characterization, such as peak and luminosity decline estimation.
- Machine learning methods help extract useful information from available data in the most efficient way.
- Several light curve approximation methods based on neural networks were considered: Multilayer Perceptrons (MLPs), Bayesian Neural Networks (BNNs), and Normalizing Flows (NFs).
- Tests using both simulated PLAsTiCC and real Zwicky Transient Facility data samples demonstrated that even few observations are enough to fit networks and achieve better approximation quality than other state-of-the-art methods.
- The authors showed that the methods described in this work have better computational complexity and work faster than Gaussian Processes.
- Using appropriate techniques increases the accuracy of peak finding and supernova classification.
- The study results were organized into a Fulu Python library available on GitHub for easy use by the community.
- This research has significant implications for future astronomical studies where large-scale surveys will provide even more data to analyze.

## Summary

The field of astronomy has been revolutionized by modern time-domain photometric surveys that collect vast amounts of data on various astronomical objects. With the advent of large-scale surveys, even more information will be available. However, most of these objects have never received a spectroscopic follow-up, which is crucial for transients like supernovae. In such cases, observed light curves could present an affordable alternative. Light curves are time series that are actively used for photometric classification and characterization, such as peak and luminosity decline estimation. However, the collected time series are multidimensional, irregularly sampled, contain outliers, and do not have well-defined systematic uncertainties. Machine learning methods help extract useful information from available data in the most efficient way. In this study, several light curve approximation methods based on neural networks were considered: Multilayer Perceptrons (MLPs), Bayesian Neural Networks (BNNs), and Normalizing Flows (NFs). These methods were used to approximate observations of a single light curve. Tests using both simulated PLAsTiCC and real Zwicky Transient Facility data samples demonstrated that even few observations are enough to fit networks and achieve better approximation quality than other state-of-the-art methods. The authors showed that the methods described in this work have better computational complexity and work faster than Gaussian Processes. They analyzed the performance of the approximation techniques aiming to fill gaps in the observations of light curves and found that using appropriate techniques increases the accuracy of peak finding and supernova classification. Moreover, they organized their study results into a Fulu Python library available on GitHub for easy use by the community. Overall, this study highlights how machine learning can help extract useful information from complex astronomical data sets efficiently. The use of neural network-based approaches provides accurate approximations for light curves with fewer observations than traditional methods while also being computationally efficient. This research has significant implications for future astronomical studies where large-scale surveys will provide even more data to analyze.