Research

galaxy zoo rings

Galactic rings are a striking and relatively common morphological feature, yet their formation mechanism and impact on galaxy evolution remains unknown.

I am currently leading a Galaxy Zoo citizen science project which aims to explore the subclasses of rings in the greatest numbers to date. In order to do this, Galaxy Zoo Mobile released 6 individual workflow campaigns, calling on thousands of citizen scientists to classify known ring galaxies into distinct morphological groups.

Using machine learning techniques, we used volunteer vote fractions to constrain the myriad of sub-classes into known and newly discovered morphological types. Results from this work are in prep and due to be published in the first quarter of 2025.

Image credit: Dawson et al. in prep

Image credit: Smirnov et al. 2024

radio transients with meerkat

Medium-timescale (minutes to hours) radio transients are a relatively unexplored population. The wide field-of-view and high instantaneous sensitivity of instruments such as MeerKAT provides an opportunity to probe this class of sources, using image-plane detection techniques.

I am currently leading the development of ML and DL techniques for mining and classifying high time cadence MeerKAT datacubes. My primary focus is on building up one branch of the team's high time cadence imaging ecosystem. This branch, named traident, is an AI focused platform for analysing MeerKAT time series datacubes. Such methods are needed owing to MeerKAT data volumes and variable noise properties.

SELF-SUPERVISED pyhsics-aware BAYESIAN NEURAL NETWORKS

I have been developing self-supervised Bayesian neural networks which can accurately infer kinematic properties from interferometric data products. Tests on simulated data products have show promising results and transfer learning tests on THINGS and WISDOM galaxies in blind test mode have yielded equally promising results. The published results for this work can be found in MNRAS as well as arXiv.

Ariel Machine Learning Competition 2019

In 2019, the ARIEL team at the UCL Centre for Exochemistry Data, hosted a machine learning competition to maximise the capabilities of ESA’s exoplanet atmospheric mission. The aim of the challenge was to remove noise from exoplanet observations caused by starspots and instrumentation.

Under the handle of SpaceMeerkat I was ranked in first place at the closing of the competition and received the prize of speaking at the ECML-PKDD conference in Wurzburg, Germany.

For the final scores of the competition and more information, please visit the ARIEL competition website.

Credit: UCL department of space and climate physics

The model used for this challenge was a 1D convolutional neural network (see image right) using the PyTorch framework and a CUDA accelerated training procedure. Much of the work that went into winning this challenge lay in the pre-processing, optimising transit curve noise cleaning, and formatting style for feeding inputs to the model.

Rather than providing our model N by 2 dimensional data comprising the transits in all channels, initial tests showed using all channels appended to one another and using a 1D convolutional network gave the highest accuracy on a short timescale.

Thanks to the use of PyTorch's integration of CUDA GPU accelerated capabilities, training the network took less than an hour on a single NVIDIA Titan xp GPU.

Credit: SARAO

Using machine learning to study the kinematics of cold gas in galaxies

In 2018 and 2019 I refined a new way of exploring the kinematics of gas in galaxies using machine learning models. These embed kinematic features into lower dimensional hyperparametric spaces and are used to rapidly differentiate disturbed gas structures from orderly rotating disks. Below is the abstract for the MNRAS accepted paper on this work which gives the results in finer detail:

Next generation interferometers, such as the Square Kilometre Array, are set to obtain vast quantities of information about the kinematics of cold gas in galaxies. Given the volume of data produced by such facilities astronomers will need fast, reliable, tools to informatively filter and classify incoming data in real time. In this paper, we use machine learning techniques with a hydrodynamical simulation training set to predict the kinematic behaviour of cold gas in galaxies and test these models on both simulated and real interferometric data. Using the power of a convolutional autoencoder we embed kinematic features, unattainable by the human eye or standard tools, into a three-dimensional space and discriminate between disturbed and regularly rotating cold gas structures. Our simple binary classifier predicts the circularity of noiseless, simulated, galaxies with a recall of 85% and performs as expected on observational CO and HI velocity maps, with a heuristic accuracy of 95%. The model output exhibits predictable behaviour when varying the level of noise added to the input data and we are able to explain the roles of all dimensions of our mapped space. Our models also allow fast predictions of input galaxies' position angles with a 1σ uncertainty range of ±17º to ±23º (for galaxies with inclinations of 82.5º to 32.5º respectively), which may be useful for initial parameterisation in kinematic modelling samplers. Machine learning models, such as the one outlined in this paper, may be adapted for SKA science usage in the near future.

For this work I used PyTorch's GPU accelerated capabilities to train a convolutional autoencoder to embed kinematic properties onto a lower dimensional manifold. From this embedding I was able to recover the circularity of gas in galaxies to a reasonably high accuracy in a fraction of the time needed for current methods while also returning information useful for seeding kinematic modelling routines.

Page updated

Google Sites

Report abuse