Galaxies are often modelled as disjoint composites of distinct spectral components, implying that different wavelength ranges are only weakly correlated. They are not. We construct a data-driven model to predict infrared emission from optical spectra, achieving almost lossless predictions. Traditional fitting methods are incapable of making predictions, being biased by model misspecifications.
Machine Learning
The biggest challenge for next-generation ground-based spectroscopic surveys is the emission from the atmosphere itself. The Subaru Night Sky Spectrograph (SuNSS) studies this emission, called airglow, which allows us to improve our calibrations, sky subtraction, and understanding of the airglow itself.
We use Graph Neural Networks to determine how galaxies relate to their environments, and determine a special neighbourhood length scale which shape galaxies the most.
We show how building a graph neural network emulator using merger trees encoded as graphs can both dramatically improve the precision with which we make simulated galaxies and improve our understanding of the galaxies we make!
This project aimed to show how building an emulator using merger trees encoded as graphs can both dramatically improve the precision with which we make simulated galaxies and improve our understanding of the galaxies we make! We achieved that with Mangrove (Paper out now)!
We go through several low - to medium level Machine Learning algorithms and optimize their performance for a classifying $V^0$ - particles, both in simulation and real data
We solve the long-standing problem of the overlapping distributions of long and short Gamma-Ray Bursts using Machine Learning
Using Graph Neural Networks, me and a group of friends showed that event reconstruction speed and precision for low-energy neutrino events can be greatly improved.