Machine learning techniques to identify ‘state-changing’ and anomalous behavior in astrophysical time series
PI: Charlotte Ward (Astronomy and Astrophysics)
The Legacy Survey of Space and Time (LSST) at Rubin Observatory is expected to discover tens of millions of transient and variable objects, from flaring supermassive black holes to variable binary star systems. The unprecedented data volumes expected from LSST have motivated the development of machine learning classifiers to classify the physical origin of light curves, such as the ALeRCE random forest light curve classifier (Sánchez-Sáez et al. 2021). Of particular interest is outlier detection: identifying previously unknown variable objects via their light curves using isolation forest algorithms (e.g. Ishida et al. 2021), temporal convolutional networks (Muthukrishna et al 2019), and autoencoder neural networks (Perez-Carrasco et al. 2023). While methods have been developed to classify transients as outliers or not, we are still lacking a key capability: quickly identifying whether a variable object has ‘changed’ its behavior within a time series. Examples include active galactic nuclei changing from typical stochastic variability to a flaring event, variable stars exploding as supernovae, or compact binaries switching from ‘high’ to ‘low’ states. In this project, we aim to identify the architectures most capable of identifying ‘state changes’ in time series, whether they be outlier detection techniques or extensions of methods that learn and predict time series behavior (e.g. latent ODEs, Sampson et al. 2024). We will experiment with different architectures on known variables and transients from the Zwicky Transient Facility (ZTF), with the goal of providing a github repo and tutorial notebooks that implement the most effective method on light curves of interest.
Expertise/skills of interest:
● Programming experience with python, julia, jax or other languages, as well as Github.
● Familiarity with machine learning architectures for classification and outlier detection e.g. neural networks, encoders.
● Useful but not required: Familiarity with time series analysis, astrophysical transients, time-domain surveys.
Expectations:
● Post-comps graduate student or postdoc with at least some experience and/or training in: (1) Astronomy & Astrophysics, Physics or a related field; and (2) Applied Math, Computer Sciences, Data Sciences, IST, Statistics or a related field.
● Write code organized with appropriate documentation that will be implemented on the ZTF/LSST light curves and released on Github.
● Weekly meeting and project updates with faculty advisor. Participate in group meetings every month.
Goal: Develop a machine learning architecture that can identify changes in variability (and therefore the underlying physics driving the variability) within time series. In particular, develop an approach to identify unusual changes to light curves of astrophysical objects from ZTF and LSST.
Specific Objectives:
● Using light curves of known ‘state-changing’ variable and transient objects from the Zwicky Transient Facility, experiment with different ML architectures to identify unusual changes in light curves. These can extend upon existing outlier classification or light curve prediction architectures designed for LSST data, or be an entirely new approach. Investigate how ‘quickly’ we can identify the new light curve behavior for different objects – ie, how many light curve data points are needed after the state change to identify the new behavior.
● Produce a public github repo that implements the classifier on light curves from ZTF/LSST.
● If time permits, write a publication on the performance of the classifier.
Engagement:
As a new faculty member looking to build connections with ICDS, Ward will participate in ICDS seminars and workshops, and will explore potential collaborations with ICDS co-hires, particularly those who are part of the Center for Astrostatistics & Astroinformatics.