Predicting Galactic Orbits with Deep Learning

Supervisor: Dr. Prashin Jethwa

Contact information: prashin.jethwa@univie.ac.at

Co-supervisors: Dr. Phillipp Petersen (mathematics dept.)

Expected duration: 9 months

Project description & Goals:

Orbit-based (aka Schwarzschild) models are amongst the most flexible dynamical models used in astronomy, making them suitable for high-quality kinematic datasets such as IFU datacubes of nearby galaxies. In the stellar-dynamics group, we develop an orbit-based model [1] which has been applied widely [2] revealing the assembly histories and dark matter distributions of galaxies. A drawback to orbit-based models is that they require numerical integration of large numbers of orbits, which can be slow to evaluate. This limits their use to relatively small samples and prohibits a complete exploration of the interesting parameter space. In this project, we will aim to develop deep-learning tools to speed-up orbit calculation, with the broader goal of making orbit-based models more widely applicable.

This is an interdisciplinary project at the intersection of astronomy, dynamics and deep-learning, and will be jointly supervised by the astrophysics and mathematics departments. There have been several recent demonstrations of the power of deep-learning to predict complex gravitational dynamics [3,4], providing timely motivation to propose this project now. The goal is that the student will learn aspects of galactic dynamics, orbit-based modelling, and will develop, train, and test a deep-learning model for orbit prediction. Candidates with enthusiasm for any or all aspects of this interdisciplinary project are welcome.

Working plan & Milestones (including final thesis):

  1. Literature and introduction to the methodology
  2. Devise an appropriate parametrization of orbits
  3. Generate training sets using existing routines for numerical orbit integration
  4. Explore orbits using statistical tools and visualisation methods
  5. Design a deep-learning architecture
  6. Evaluate performance of the deep-learning model
  7. (Optional) Iterate steps 5-6 to improve performance
  8. Write thesis

References: