Research
I am interested in understanding how galaxies grow and evolve. My research program focuses on galaxy evolution in the context of their surroundings, ranging from low-mass satellite galaxies in the low-redshift universe to massive galaxies residing in enormous clusters. By using wide-area astronomical surveys, and simulation data, targeted multi-wavelength observations, and interpretable machine learning (ML) techniques, we can investigate the interplay of galaxies’ internal physical processes and environmental effects.
Interpretable ML in Astrophysics
Standard deep learning models are remarkably effective at predicting galaxy properties from images, but their internal workings often remain opaque. Learned features can be distributed across many neurons (superposition), and single neurons can represent multiple concepts (polysemanticity), obscuring the link between galaxy appearances and physical properties.
In an effort to remedy this, I’ve developed Sparse Feature Networks (SFNets). These networks learn a large dictionary of distinct morphological features but activate only a very small number for any given galaxy. Physical properties can then be estimated as a simple linear combination of these sparse, active features. SFNets provide direct interpretability: we can see exactly which learned image features contribute to a prediction, without sacrificing the predictive accuracy achieved by deep neural networks. By leveraging mechanistic interpretability techniques directly during the model training procedure, SFNets enable us to build a clearer morphological and physical understanding of galaxy evolution.
Graph Neural Networks
Graph neural networks, or GNNs, are remarkably adept at representing galaxies and their physical interactions admist large scale structure. My collaborators and I have successfully used GNNs to augment the galaxy-halo connection by incorporating information from galaxy surroundings. For “painting” galaxies onto dark matter halos, we found that GNNs outperform abundance matching and other ML methods, likely because they can model the galaxy-halo-environment connection on >3 Mpc scales.
Convolutional Neural Networks
Convolutional neural networks, or CNNs, can represent complex morphological information in images. Because galaxy appearances encode their physical properties and formation history, we can train CNNs to extract information that might otherwise be only accesible via spectroscopic observations or detailed physical modeling. For example, we are able to predict galaxy optical spectra directly from image cutouts.
Using these techniques, we can also estimate galaxies’ gas metallicities, neutral hydrogen gas mass, nuclear activity, and excitation state. We have also found CNNs to be useful for enhancing X-ray observations and removing artifacts from photometric catalogs.
Low-mass Galaxies
Low-mass galaxies are vital for studying galaxy formation physics, because they are so sensitive to feedback processes and are easily disrupted by interactions with other (more massive) galaxies. In the very local Universe (z < 0.03) we are discovering many new low-mass galaxy candidates with machine learning, and targeting them for spectroscopic confirmation as part of a DESI survey.
Gas in Diverse Galaxy Populations
As part of my thesis work, I studied the star-forming, multi-phase, interstellar contents of all kinds of galaxies. These galaxy populations include members of massive clusters, Lyman break galaxy analogs, and gas-rich galaxies at z > 1.