Research
Time-domain astronomy and sky surveys
Large surveys that image the sky night after night, and the machine learning that turns their alert streams into classified, followable events.
Nancy Grace Roman Space Telescope (RAPID)
RAPID is the Caltech-IPAC pipeline that will find time-variable sources in Roman data. I lead the machine learning effort.
Zwicky Transient Facility (ZTF)
ZTF is a time-domain survey with a 47 square degree field of view, on the Samuel Oschin 48-inch Schmidt telescope at Palomar Observatory. I lead the machine learning for the survey: separating real sources from bogus ones, finding streaking asteroids, and sorting variables and transients into their classes. Classical methods for small data sets, deep learning when labels are plentiful.
Vera C. Rubin Observatory / LSST
Rubin's Legacy Survey of Space and Time is a ten-year survey from an 8.4 m telescope in Chile, about a thousand visits per field in six filters. I was a co-chair of the Transient and Variable Stars Science Collaboration for a few years and am part of the Informatics and Statistics collaboration.
ALeRCE alert broker
ALeRCE (Automatic Learning for the Rapid Classification of Events) is a Chile-led broker that classifies alerts, trained on the ZTF stream and built for Rubin. I contributed labeled data from CRTS and ZTF and take part in the classification work, including teaching at the La Serena School for Data Science.
AI/ML methods and measurement
Methods that are not tied to one data set: interpretable deep learning, classification under sparse and uneven sampling, and how to measure what a model is doing.
Explainable AI for gravitational-wave alerts (GWSkyNet)
GWSkyNet and GWSkyNet-Multi classify LIGO-Virgo-KAGRA alerts within minutes. The focus is on understanding what the networks have learned.
NIST
Text to come.
Methodology transfer to the life sciences
The same problems show up in medicine and biology: few labels, noisy measurements, data spread over many sites. Tools from astronomy carry over.
Fungal contaminants on space missions
Predicting which fungi could contaminate spacecraft, by screening whole proteomes for orthologs of proteins linked to survival in harsh conditions. Preprint on bioRxiv, 2026.
Early Detection Research Network (EDRN)
EDRN is a National Cancer Institute network of tens of institutions that moves cancer biomarker research toward clinical use. With the Data Science group at JPL I work on machine learning for early detection and on building the data sets it needs. Domain experts stay in the loop, but once the data are abstracted, methods from astronomy apply with little change.
Biomarker knowledge bases (BiomarkerKB, CFDE)
BiomarkerKB brings biomarker data from many sources into one knowledge base that can be explored by biomarker. It is part of the NIH Common Fund Data Ecosystem (CFDE).
Dormant
Fast radio burst morphology
Deep learning to classify fast radio bursts by the shape of their dynamic spectra.
Earlier work
Catalina Real-time Transient Survey (CRTS)
CRTS used the Catalina Sky Survey asteroid-search data to find rare transients and variables over thirty thousand square degrees. I worked on transient classification and built parts of the early infrastructure.
RoboPol
RoboPol is a four-channel optical photopolarimeter on the 1.3 m telescope at Skinakas Observatory, Crete. It measures linear polarization in a single exposure. Its main program monitored more than a hundred gamma-ray bright blazars to test models of jet structure and magnetic fields. Partners: Caltech, MPIfR Bonn, Nicolaus Copernicus University, IUCAA, University of Crete.
Molecular and Cellular Characterization of Screen-Detected Lesions (MCL)
MCL teams study tumor tissue, cells and microenvironment to tell screen-detected early lesions apart from interval and symptom-detected cancers. The work overlaps with EDRN. Our virtual reality view of tumor data was shown at SIGGRAPH 2019.
Tremor classification from hand-drawn spirals
A deep learning classifier for tremor types from spirals drawn by patients (Movement Disorders, 2025).
VIFI (NSF DIBBs)
A Virtual Information-Fabric Infrastructure for analysing distributed, fragmented data sets without moving them and without exposing sensitive raw data. Analyses run where the data live, in secure containers, with search and metadata built on persistent identifiers.