Research
Discovering electromagentic counterparts to gravitational waves
With LIGO-Virgo-Kagra's next observing run (O4) on the horizon, there is a need to reevaluate human-centered protocols for finding and following up kilonovae, as data throughputs have overwhelmed the ability to manually synthesize alerts for devising and coordinating necessary follow-up with limited resources.
Pythia is an AI agent that strategizes how to allocate a limited follow-up budget to identify and follow up kilonovae (KNe) among several tens of contaminant sources within the gravitational wave localization region. It trains on simulated scenarios given gravitational wave triggers during O4 to evaluate millions of strategies and evaluate the optimal sequence of follow-up decisions given a scenario. It uses reinforcement learning to evaluate the explore-exploit tradeoff, solve the credit assignment problem given delayed information from actions chosen, and compute the optimal sequence of decisions under uncertainty.
Pythia is presented 9 transient light curves from the Zwicky Transient Facility (ZTF), one of which is the KN and the rest are contaminants, chosen randomly from a list of supernovae and unassociated GRB afterglows. The agent observes the events on day 1 and on days 2 through 7 assigns one additional photometry with ZTF in g, r, or i using a deep 300s exposure to one of the events. The resulting forced photometry or upper limit is added in the next timestep. The agent gets a reward 1 if the follow-up is assigned to the KN, and 0 otherwise. The objective is to maximize the number of follow-up assigned to the true KN. Pythia is 3.5x better than random (Sravan et al., submitted).
Random sample of ZTF SN Ia LCs augmented in real-time given the goal of minimizing SALT2 parameter uncertainties
Improving distance estimates using Type Ia supernovae
Even as surveys have produced an unprecedented sample of thousands of spectroscopically confirmed SNe Ia, driving down statistical uncertainties, systematic uncertainties remain a major roadblock toward precision cosmology. As LSST promises to plough deeper at the expense of temporal resolution, it is important to have light-curves that maximize information for cosmology.
We design a novel framework for performing real-time science-driven follow-up and demonstrate it for strategizing photometric augmentation of Zwicky Transient Facility (ZTF) SNe Ia light-curves given the goal of minimizing SALT2 parameter uncertainties. We find a median improvement of 2-6% for SALT2 parameters and 3-11% for photometric redshift with 2-7 additional data points in g, r and/or i compared to random augmentation (Sravan et al. 2021).
Recommender Engine For Intelligent Transient Tracking
While next-generation surveys like the Rubin Observatory (https://www.lsst.org/), monitoring the sky with unprecedented spatial and temporal resolution, are leading to transformative advances in our scientific understanding, the resulting data deluge have rendered traditional human-guided data collection and inference techniques impractical. In order to maximize the science potential of data collected and limited follow-up resources, autonomous systems reacting in real-time to simultaneously maximize diverse science goals are needed.
REFITT is a novel AI system that strategizes and coordinates transient follow-up in real-time, given the goal of most optimally augmenting the alert stream for science inference. Right: Follow-up recommendations on a random night of Rubin Observatory operations given 32,000 core-collapse supernova events of interest (Sravan et al. 2020).
Single and Binary Stripped-Envelope Supernova Progenitors
The mechanisms driving removal of envelopes of stripped-envelope supernova progenitors is a key challenge to our understanding of massive star evolution. Comparing population-scale detailed stellar evolution models with a wide variety of observational constraints (serendipitous progenitor photometry, multi-band SN light-curves, X-ray/radio observations) not only constrain evolutionary channels but can also reveal unobserved populations.
Right: Probability distribution of progenitor and companion properties of SN 2016gkg (Sravan et al. 2018).
Center of Expansion and Age for Supernova Remnants
By tracking the proper motion of supernova ejecta over several years and rewinding their trajectories, I estimate the spatial location and date of the supernova explosion. More recently, I have been involved in developing image segmentation-localization-tracking techniques to follow the motion of large numbers of knots in high-resolution HST images and obtain improved estimates with well-characterized systematics.
Credit: STScI Press Office