Research

Building deep learning tools to optimize the Legacy Survey of Space and Time (LSST) telescope in real-time. The main challenge is training neural networks on synthetic data, then getting them to work on actual telescope images where conditions are never quite what you expect.

I'm working on compression algorithms that can reduce data volumes by 15+ times while keeping all the important astronomical information intact. This is crucial because LSST will generate about 20TB of data every night for 10 years.

Developed a new way to model how neutral hydrogen is distributed in the universe after reionization. The traditional approach assumes a simple power-law, but that misses a lot of the physics happening in self-shielding gas clouds.

I created a three-step neural network that learns from high-resolution simulations and can predict gas properties much faster than running full simulations. This lets us explore parameter space efficiently and compare different reionization scenarios against observations.

Building a Julia package to find stable periodic orbits in multi-planet systems. The code uses automatic differentiation to compute fast, accurate gradients and Levenberg-Marquardt optimization to search for orbits that repeat exactly.

This is particularly useful for analyzing exoplanet systems where you want to know if the observed orbital configuration is actually stable over long timescales. The package also supports transit timing constraints.

Applied wavelet scattering transforms to quasar absorption spectra to extract cosmological information that traditional power spectrum analysis misses. The idea is that these "cosmic barcodes" contain non-Gaussian information about dark matter that we're not currently using.

Showed that this approach can improve cosmological parameter constraints by an order of magnitude compared to standard methods. This could be crucial for upcoming surveys like DESI that will measure millions of these spectra.

Used data from the cosmic microwave background to test whether fundamental constants like the fine-structure constant have changed over the 13.8 billion year history of the universe. This was my undergraduate thesis project.

We didn't find evidence that constants are changing (which is probably good for physics!), but we set some of the tightest constraints to date using two different theoretical models.

Worked on Project Premonition, which aims to track disease vectors by analyzing blood collected by mosquitoes. My job was to teach neural networks to identify mosquito species from the sound of their wingbeats.

Built a time-series LSTM that significantly outperformed the existing approach. The main challenge was filtering out environmental noise from the actual species-specific wingbeat patterns.

Galaxy clusters should be cooling catastrophically, but they're not. The leading explanation involves heating from supermassive black holes that create "bubbles" visible in X-ray images.

I ran simulations to test whether the correlations we see in observations are actually specific to black hole feedback, or just generic properties of hot cluster gas. Turns out some of what we thought were real cavities might just be noise in the data.

Used the AdS/CFT correspondence to compute correlation functions for weird quantum field theories called Lifshitz theories. The cool thing about this approach is that hard quantum field theory calculations become easier gravity problems in higher dimensions.

Also worked on plasma physics, using something called bispectral analysis to understand energy transfer in plasma systems. That code is now being used to study spiral arm stability in galaxies.