Research

Causal foundation models: separating physics from instrument

Every measurement of the physical world mixes the signal of interest with artifacts of the sensor that recorded it. This confounding limits what we can learn about the underlying physics and makes it hard to combine data across instruments. I develop representation learning methods that explicitly disentangle the two: a dual-encoder architecture with a counterfactual generative objective that uses overlapping observations of the same object from different instruments (Learning What’s Real, NeurIPS 2026), and causal foundation models for stellar light curves that factorise a star’s physics from the properties of the telescope that observed it (ICML 2025). The resulting representations enable counterfactual view generation, parameter inference and instrument-independent similarity search, and they provide a general recipe for scientific self-supervised pretraining.

With MIT’s Generative AI Impact Consortium I co-lead a project to build a time-series foundation model for astrophysics, and with collaborators across physics we released PhyTS (NeurIPS 2026), the first unified benchmark for scientific time series spanning gravitational-wave detection, dark matter searches, neutrino mass measurement and exoplanet discovery, across twelve orders of magnitude in sampling rate.

Generative models of instrument systematics

Space telescopes produce structured noise, such as scattered light from the Earth and Moon in NASA TESS images, that cannot be observed in isolation and therefore cannot be removed by supervised denoising. BEAM (Background Estimation via Additive Modeling) learns a generative model of the clean signal and uses it to score candidate noise estimates by how plausible the underlying image looks, turning score matching into a tractable likelihood sampled with Langevin dynamics. A conditional flow model also forecasts the stray light from spacecraft, Earth and Moon geometry alone, before an observation is taken.

BEAM produces cleaner stray-light estimates than the filtering tool used in current TESS pipelines, and its blind forecast of scattered light across the observing window of the interstellar comet 3I/ATLAS was used to position the spacecraft for TESS’s January 2026 observations of the comet. The work is funded by a NASA TESS Guest Investigator award and opens a path to recovering faint signals that current pipelines discard.

Real-time discovery in the transient sky

Surveys such as the Vera C. Rubin Observatory’s LSST and the Zwicky Transient Facility issue millions of alerts each night, and the most interesting objects must be identified while they are still evolving. During my PhD I built RAPID, a recurrent neural network that classifies explosive transients from the first few nights of photometry (250+ citations; integrated into the Fink and Lasair alert brokers), DASH, a deep learning classifier that reduced supernova spectral classification from days to seconds, and the first real-time anomaly detection framework for large time-domain surveys, which combines Bayesian models and temporal convolutional networks (MNRAS 2022, Paul Murdin Prize).

More recently my students and I developed MCIF, a classifier-based approach to multi-class anomaly detection (RASTI 2024), showed that simulation-based pretraining reaches fully supervised performance with an order of magnitude fewer labels and transfers from ZTF to LSST (MNRAS Letters 2025; ICML 2025), and deployed the resulting classifier and anomaly detector live on the UK-led Lasair broker for every LSST alert. Related work predicts the age of a transient in real time from multivariate time series (NeurIPS 2023; featured in New Scientist), estimates supernova ages for the Young Supernova Experiment with Bayesian methods, and extends DASH into AstroDASH 2.0, a web platform and REST API built with the SkAI Institute.

Exoplanets and NASA’s TESS mission

I lead the Astronet group for NASA’s Transiting Exoplanet Survey Satellite, coordinating twelve software engineers, research scientists, faculty and PhD students, including volunteers from Google and Meta, to deliver the production deep learning pipeline that vets planet candidates in TESS data. The pipeline has contributed to hundreds of candidate discoveries; its latest generation, AstroNet-Vetting-v2 (AJ 2026), automates exoplanet discovery in TESS full-frame images. I also work on the TESS Quick-Look Pipeline, on stellar variability classification and on unusual systems such as a disintegrating rocky planet with comet-like tails (ApJL 2025), and with the Center for Astrophysics on robust detection of biosignature gases in exoplanet atmospheres (NeurIPS 2026 ML4PS workshop).

Surveys, collaborations and community challenges

I am a long-term member of the LSST Dark Energy Science Collaboration, where I co-developed the Photometric LSST Astronomical Time-series Classification Challenge (PLAsTiCC), a community-wide classification challenge joined by more than 1,000 teams on Kaggle, and of the Young Supernova Experiment and the Habitable Worlds Observatory AI/ML Working Group. Earlier I worked with the Dark Energy Survey and OzDES on supernova cosmology, spectroscopic classification and data releases, with 4MOST’s Time-Domain Extragalactic Survey, and with the SkyMapper Transient Survey, where I helped run the BBC Stargazing Live citizen-science search for Planet 9. I co-organised the Boston Astrophysics x Machine Learning Hackathon (2024), the Accelerating Physics with ML workshop at MIT (2023) and the TESS Science Conference III (2024).