Classifying Weakly Detected Gamma-ray Transients
PI: James DeLaunay (Astronomy and Astrophysics)
Department of astronomy and astrophysics has promised to fund the tuition of the graduate student.
Project Description:
This project is meant to be accessible to those with no astronomy background. The PI and the group will assist in teaching any necessary astro specific concepts and aid in using astro specific tools and data.
Gamma-ray bursts (GRBs) are brief flashes of gamma-rays lasting from milliseconds to minutes and are the brightest phenomena in the Universe. There is a bimodal distribution of GRB durations, where the longer ones are believed to originate from the collapse of massive stars and the shorter ones are believed to originate from two incredibly dense stellar remnants (neutron stars) colliding into each other at near the speed of light. The existence of these two populations in the observables are very apparent, but there is significant overlap, making classification non-trivial. Additionally, there are several other astrophysical transients at similar photon energies and similar durations to GRBs that are regularly detected by GRB monitors, such as solar flares and outbursts from highly magnetized neutron stars (magnetars) in our galaxy or other nearby galaxies. Rapid, real-time classification of GRB-like signals are essential to inform the strategy on how to further study these events with other telescopes. Classification is also essential for statistical studies of these populations.
There has been recent interest in creating more sensitive searches for GRBs in order to better study less luminous bursts. Most specifically after the 2017 joint detection of a very low luminosity GRB and gravitational waves from a merger of neutron stars. This has resulted in the development of search pipelines that are more computationally intensive but much more sensitive. These pipelines have been very successful in discovering more GRBs, but since these additional discoveries are all weak signals they are even more difficult to classify.
GRB monitor data is a time series with each entry being the number of photons detected in a short time span with several channels (bins of measured photon energy and several detectors). GRB monitors have high background rates, detecting hundreds to thousands of photons a second and rely on the short duration of GRBs to be able to detect them above the background and use times before and after the GRB to estimate the background rate. Using the response of the detectors the shape of the distribution of photon energies (spectrum) is estimated.
This project will consist of finding the optimal way to perform classification on these weakly detected gamma-ray transients, by exploring different
● AI techniques
● Feature inputs (calculated properties, raw time series data, or both)
● Training data (sample of hundreds to thousands of detections in real data or could use simulation for more statistics)
This project will use data from the Fermi Gamma-ray Space Telescope and the Neil Gehrels Swift Observatory. The Swift satellite has its mission operation control center housed here at Penn State!
Level of Effort:
Graduate student, 1 year at 25% RA
Desired Skills:
● Expertise in programing language with extensive machine learning packages (preferably Python)
● Experience with using and training various machine learning models that may be relevant to the project (such as: support vector machines, nearest neighbor algorithms, convolutional and recurrent neural nets, etc.)
Project Objectives:
● Find and train an optimal AI model to classify weakly detected gamma-ray transients
● Integrate classifier into current analysis pipeline to classify in real-time
● Provide thorough documentation in how the training was performed and how to use the classifier, so future group members can easily use it or update the model if needed
● Attend weekly group meetings and meetings with PI
Medium to Long term Goals:
● Medium – provide classification probabilities in real-time alerts that we broadcast to astronomers around the world when we detect a new transient
● Long term – We are at the start of a three year funded project to run these sensitive searches over the full history of Fermi (16 years) and Swift (20 years) data. In doing so we will discover over a thousand new gamma-ray transients that will need classification to be published in a catalog. We will also need classifications to perform population studies on the different types of GRBs.
PI’s engagement with ICDS:
PI is a heavy user of the roar collab cluster and has taken part in programs to provide testing and feedback. PI will start regularly attending ICDS activities.