Research

When LSST begins its 10-year survey, one of the most ambitious scientific instruments ever built will need to maintain optical alignment every 36 seconds as it scans the sky. I developed AIDonut, the first neural network deployed for real-time control of telescope optics, now running in the Rubin Active Optics System. From out-of-focus "donut" images it predicts 21 Zernike coefficients plus seeing, reducing wavefront estimation error by 20% relative to the forward-modeling baseline solver and cutting inference from 30–40 seconds to under one. That meets the 36-second visit cadence and shortens the correction latency from three visits to two.

The core challenge was sim-to-real transfer: the network was trained on physics simulations, but real telescopes are messy — atmospheric turbulence, mechanical vibrations, thermal deformations. Adapting it to on-sky LSSTCam data across 3 million donut stamps, I tested five transfer-learning strategies and redesigned the network with per-mode output heads and a bounded GradNorm loss. The system deployed to production in September 2025.

Before an LSSTCam image can be used for science, the raw detector data passes through a nonlinear sequence of calibration steps. I built a convolutional surrogate that replaces the whole sequence with a single 282 ms evaluation. It reproduces fluxes to 0.1%, against 0.7% for a linear gain-bias-flat model, and reduces noise by 59% compared to a linear model, evidence that it recovers the nonlinear terms in the detector response.

I'm now extending it to full-detector scale, including crosstalk between amplifiers. The goal is an end-to-end path from raw amplifier data to wavefront correction, fast enough to run within the camera readout.

KBMOD searches for faint moving objects (distant trans-Neptunian objects) by shifting and stacking images along candidate trajectories. The existing classifier worked well on bright detections but failed on the fainter 5σ candidates where the real discoveries live. I found the previous model had never learned any astronomy at all; it had defaulted to a trivial pixel threshold that gave perfect accuracy on high-SNR data.

I fixed the data pipeline and designed a ResNet with squeeze-and-excitation channel attention, trained with focal loss on 2 million samples, so the network can learn which information matters at different signal-to-noise levels. It reaches 97% detection accuracy on a held-out test set and recovered 3,318 novel high-confidence candidates that aren't in existing cross-match catalogs, potential new trans-Neptunian objects. We are now fitting orbits to confirm them.

A new project developing a multi-scale foundation model built around how information clusters across different physical scales.

Astronomy papers carry much of their content in figures, but literature search indexes only text, so there has been no way to find a plot by describing what it shows. FOTO is a figure-level search tool over 482,750 captions from 53,839 peer-reviewed astro-ph papers. Each figure is indexed once by its title and caption using a 110M-parameter open embedding model that runs locally, and a large language model is applied only to the retrieved shortlist, where it verifies real figures rather than inventing references to ones that don't exist. On a multi-register benchmark, held to the stricter standard of finding the exact figure, FOTO places the target in its top 20 for 29–79% of queries, against 6–20% for Pathfinder and 8–16% for Semantic Scholar at simply finding the right paper.

Sometimes you remember the plot but not the paper it came from. Describe what you remember, and FOTO will find the figure so you can cite the right paper.

Cosmological simulations produce more data than we can afford to keep at full precision. I developed a lossy neural compression method that quantizes each field on a fixed grid, so the reconstruction error is set by the science tolerance rather than by training, and learns only the bin probabilities with a two-scale autoregressive transformer.

Benchmarked on weak lensing maps, an N-body density field, and Lagrangian displacements, it reaches percent-level accuracy at 1.3–5.3 bits per pixel (versus 32 stored) while preserving the power spectrum to better than 1%. It achieved the lowest rate on every field tested, 9–27% below a Gaussian coder and 6–21% below SZ3. That gap is a measurement in itself: it's the structure that two-point statistics and local smoothness can't represent.

Standard exoplanet detection methods can't uniquely determine both mass and orbital eccentricity — you get a family of degenerate solutions. Periodic orbits, configurations where a planetary system returns to its exact initial state, exist at discrete points in parameter space and break this degeneracy. I co-developed a Julia package that combines periodic orbit constraints with transit timing variations, using a Levenberg-Marquardt optimizer with Jacobians computed by automatic differentiation that converges in hundreds of iterations. We validated it by reproducing HR 8799's 4-planet periodic orbit (stable over 10,000 years), then applied it to TRAPPIST-1, where it found evidence for a previously undetected planet.

That result led to an accepted JWST Cycle 4 program (#12492), on which I'm a Co-I, to confirm the candidate planet with follow-up observations.

Understanding how the first stars ionized the universe requires simulations that resolve kiloparsec-scale gas clumping inside hundred-megaparsec cosmic volumes, a computational problem that's fundamentally intractable at full resolution. I built a deep-learning emulator, a residual multi-layer perceptron trained on 2 kpc-resolution radiation-hydrodynamic simulations spanning 126 parameter combinations. It takes simulation parameters (redshift, radiation field strength, local density, reionization timing) as input and directly predicts the ionizing photon mean free path, with no new simulation needed.

The emulator achieves 1.6% median error across nearly four orders of magnitude in mean free path and runs in milliseconds on standard hardware. I used it to constrain the reionization history from observed mean free path measurements at z = 5–6, which is consistent with Planck CMB measurements but derived from completely independent observations. The paper also reveals a 2–3× decline in ionizing emissivity that standard power-law models of mean free path evolution can't explain.

The standard approach in Lyman-α forest cosmology uses power spectra — essentially just the Fourier transform squared. It captures Gaussian information but throws away all phase relationships and non-Gaussian structure. Everyone knew this was lossy, but quantifying the loss was hard. I applied the wavelet scattering transform — a technique from signal processing that preserves both second-order statistics and higher-order moments through a cascade of wavelet transforms and nonlinearities — and showed it improves cosmological parameter constraints by orders of magnitude using Fisher information analysis.

The deeper insight: representation matters more than algorithm. The right mathematical encoding of the data unlocked information that was always there but invisible to traditional methods. This idea — that finding where information lives is the real bottleneck — has shaped everything I've done since.

My undergraduate thesis. Used Planck CMB data to test whether the fine-structure constant has changed over 13.8 billion years by solving the dynamics of scalar field models and constraining their free parameters with principal component analysis and Monte Carlo simulations. We didn't find evidence of variation — probably good news for physics — but set some of the tightest constraints to date.

Worked on Project Premonition, which tracks disease vectors by analyzing blood collected by mosquitoes. Built a time-series LSTM to identify mosquito species from wingbeat interference pattern, outperforming the existing baseline. The main challenge was separating species-specific wingbeat patterns from environmental noise which is a signal processing problem that connected nicely to my later work on information extraction.

Galaxy clusters should be cooling catastrophically, but they're not — supermassive black holes create "bubbles" that heat the gas. I ran simulations testing whether the correlations we observe are actually specific to black hole feedback or just generic properties of hot cluster gas. Some of what we thought were real cavities might just be noise.

Used AdS/CFT correspondence to compute correlation functions for Lifshitz field theories — hard quantum field theory calculations that become tractable gravity problems in higher dimensions. Separately, wrote bispectral analysis code for studying energy transfer in plasmas, now being used to study spiral arm stability in galaxies.