Improving the data-analysis pipeline of the COSI space mission with machine learning and more
Andreas Zoglauer, Staff Researcher
Space Sciences Laboratory
Applications for Spring 2024 are closed for this project.
COSI, the Compton Spectrometer and Imager, is a NASA-funded gamma-ray telescope which is currently under development and scheduled for launch in 2027. It will observe Galactic nucleosynthesis and positron annihilation, as well as the most violent events in our Universe (supernovae, neutron star mergers) and the most extreme environments (pulsars, black holes).
This project is centered on improving the data-analysis pipeline of COSI by applying the latest machine-learning and other methods to individual segments of the pipeline. The available topics are:
(1) Approximate the imaging response using a deep neural network. This includes Monte-Carlo simulations of the instrument, improving an existing neural network to approximate the imaging response, and implementing a Richardson-Lucy image deconvolution approach to test the imaging response to create images.
(2) As an interation of the above, try to improve the memory footprint of the existing, binned response.
(3) Improve the Compton event reconstruction, i.e., determine the Compton-scatter path of the gamma rays in the detector via machine learning, and work on improving the verification pipeline to test the machine learning approaches.
(4) Work on implementing an energy loss approach and an electron tracking approach for a successor mission of COSI (launch in ~15 years) using machine learning.
All topics are novel and should - if we get it to work - ultimately lead to published papers.
Role: We will create URAP teams consisting of 2-4 students. Each team will work on one topic. The ultimate goal is to improve upon the currently existing reference implementations with new, cutting-edge approaches. The team will meet weekly for 1 hour with your advisor via zoom, but we also expect the URAP team members to meet among themselves regularly to work on the project together. Data sets containing the test and training data will either be provided or can be easily created via simulations/
The concrete learning goals depend on your chosen topic. In general, they include:
(1) Learn about the data analysis of a modern NASA-funded telescope, COSI
(2) Learn how to use state-of-art analysis tools (e.g. for machine learning topics TensorFlow/Keras) or learn hwo to develop your own.
(3) Learn how to apply these toolkits to the data of a modern telescope
Since this is a multi-year project, we prefer students who want to work with us for more than 1 semester. The time spent on the project should be on average at least 6 hours per week, but more is strongly preferred.
During the semester, we will predominantly meet and communicate via Zoom, Slack, and email.
Day-to-day supervisor for this project is Dr. Andreas Zoglauer who is the COSI project scientist and data pipeline lead.
In your application, please clearly indicated the following:
(1) Which project(s) you are interested in (COSI, GAPS or both) and in case of COSI which topics you are interested in (imaging response, event reconstruction, energy loss, electron tracking) - please rank them
(2) Why you have applied for this position, especially, why you are interested in doing a data-science project related to astrophysics
(3) Your previous experience with python, git, and/or bash
(4) Your previous experience with Linux and remote code execution via ssh
(5) Your previous experience with machine-learning tool kits
(6) How many hours per week you can spend on the project
(7) Assuming you like the project, if you would be interested to work more than one semester on the project.
Qualifications: We will assemble teams of 2-4 students with different experience levels and backgrounds. Thus while the following qualifications apply to seniors, consider these qualifications relaxed for first-year students. The desired qualifications are: Interest in Physics and Astrophysics; Data 8 or equivalent; proficiency in Python or C++; prior experience with machine learning (e.g. (deep) neural networks, (random forests of) (boosted) decision trees, support vector machines) and with a common machine learning library recommended (e.g. tensorflow, keras, TMVA); familiarity with the Linux/Unix environment, bash, and git; familiarity with remote code execution via ssh; great organization and communication skills; punctuality and reliability; talent for multitasking and balancing this project with your normal classes.
Hours: 9-11 hrs
Off-Campus Research Site: We will mostly communicate via zoom. However, the URAP teams should meet once per week in person on campus.
Related website: http://cosi.ssl.berkeley.edu
Related website: https://github.com/zoglauer/gamma-ai
Engineering, Design & Technologies, Mathematical and Physical Sciences, Digital Humanities and Data Science