Predictability of bursts of a recurrent nova using topological data analysis and machine learning

TL;DR

This paper combines topological data analysis and machine learning to predict bursts in the recurrent nova RS Oph within a year, demonstrating high accuracy with persistence landscapes features.

Contribution

It introduces a novel approach using persistent homology features for predicting nova bursts, achieving reliable results with supervised learning.

Findings

High recall and accuracy in predicting nova bursts within a year.

Persistence landscapes effectively capture relevant features for classification.

Method demonstrates potential for astronomical event prediction.

Abstract

RS Oph is a recurrent nova, a kind of cataclismic variable that shows bursts in a period approximately shorter than a century. Persistent homology, a technique from topological data analysis, studies the evolution of topological features of a simplicial complex composed of the data points or an embedding of them, as some distance parameter is varied. For this work I trained a supervised learning model based on several featurizations, namely persistence landscapes, Carlsson coordinates, persistent images, and template functions, of the persistence diagrams of sections of the lightcurve of RS Oph. A tenfold cross validation of the model based on one of the featurizations, persistence landscapes, consistently shows high recalls and accuracies. This method serves the purpose of predicting whether RS Oph is bursting within a year.