I fit bad models to worse data.
I'm an Australian Research Council Discovery Early Career Researcher Award (DECRA) fellow at Monash University, in the School of Physics and Astronomy.
My current research in astrophysics focusses on stars that should not exist. Astrophysics can explain the properties of 99% of stars in the universe. However, around 1% of stars cannot be explained given our current understanding of how stars form or how chemical elements are created. The very existence of these stars represents some of the most significant gaps in our knowledge of stellar astrophysics. I'm searching large astrophysics databases to find enough of these kinds of stars so that I can ultimately explain their origin.
I also work on projects with a focus on applied machine learning, that are unrelated to physics.
Scroll down to student projects or resources, which includes my research expectations for current and prospective students.
1As of 2019 I am currently five years post-PhD. I have published over 100 peer-reviewed publications that have collectively accrued more than 5,400 citations. My current h-index
NASA's Astrophysics Data Service (ADS) is commonly regarded as the most reliable database of publications and citations within astrophysics. But there are plenty of ways to search my research outputs:
Summaries and additional content from recent publications will be placed here.
2I am currently offering undergraduate research projects (e.g., PHS2250, PHS3350), honours-level projects, and PhD theses. Below is an incomplete list of the projects I am currently offering.
Contact me if you are interested in any of these projects.
Designing the right architecture for a deep neural network can be an art. There's no good theory that motivates network design.
Weight agnostic neural networks are a new idea in machine learning that allows a neural network to simultaneously train the weight coefficients, and the architecture of the network itself. Ideas from information theory are used to help guide the location of new nodes. Although this approach is new, studies to date have shown that the resulting network can have orders of magnitude fewer nodes, making them less likely to overfit and be more susceptible against 'hot pixel attacks'.
In this project you will devleop and train weight agnostic neural networks with applications to physics problems.
Ultra metal-poor stars are those with 1/10,000th (or less) of relative iron contant than what is found in the Sun. These objects formed soon after the Big Bang, and are local relics of star formation in the high-redshift universe.
In this project you will use data-driven methods to identify previously undiscovered ultra metal-poor stars in the fifth data release of LAMOST.
We find planets around stars nearly everywhere we look. When we don't find planets, we can only place limits on how many planets there can be based on the limits of our method and detector. The most common technique to find exoplanets is through radial velocity monitoring.
Radial velocities aren't informative if the exoplanet system is face-on towards us. But astrometry is sensitive to exoplanets, irrespective of their orientation with respect to us.
The Gaia space telescope recently released astrometry for over one billion stars in the Milky Way. In this project you will use the astrometric noise, self-calibrated from many Gaia sources, to estimate how much astrometric noise a star should have given it's properties. By combining these results with simulations, you will estimate limits on the total mass in planets that nearby star systems can contain. This will be the first ever candidate list of star where there is growing evidence that they host no planets, an important perspective for understanding why stars do have planets.
Hg-Mn stars are among the most perplexing kinds of stars known. They appear simple, like a textbook case of the competing effects between gravity and radiative pressure. But for two Hg-Mn stars of the same mass, luminosity, metallicity, and colour, their detailed chemical abundances can vary by orders of magnitude. This variation cannot be explained by radiative diffusion or gravitational settling.
They are main-sequence stars (i.e., so they have not evolved into red giants) that do not rotate quickly and do not have detectable magnetic fields. How can two identical stars show such variations?
In this project you will perform a comprehensive review of Hg-Mn stars with a focus on estimating their astrophysical parameters and chemical abundances (from archival spectra) in a self-consistent way, and test specific hypotheses that could explain their origin.
The Anglo-Australian Telescope is the largest optical telescope in Australia. In the past few decades it has collected some 25 million spectra of stars and galaxies. These data are all publicly accessible, but they are in their raw format. Those data need to be processed from their raw format to a reduced or calibrated format in order for them to be scientifically useful. Normally astronomers would do this 'data reduction' process interactively, or by closely watching an automatic process for failures.
In this project you will extend the existing automatic calibration software with a focus on including various machine learning techniques at every step to help monitor the processing of all data. This will include supervised and unsupervised machine learning techniques, with a focus on deliverable outputs (e.g., reduced data products or autonomatically-generated reasoning why procedures failed).
Experience in programming is essential to this project. No machine learning experience or astronomy background is necessary for this project, but it would be advantageous.
Sentinel-2 is a European Space Agency mission that provides high-resolution images of Earth every fortnight or so. These images are publicly accessible.
Motivated by the fact that the world is on fire, and oil companies have strong vested interests in under-reporting spills, in this project you will use satellite imagery to identify and monitor oil platforms for oil spills.
This project will build upon or require expertise in processing large volumes of data, and applied machine learning techniques.
If you are interested in doing a research project but you are not sold on any of the projects above, just contact me.
3