Resolving galaxies across cosmic time
JWST now resolves galaxies at kiloparsec scales out to z ~ 8, and Euclid, Rubin/LSST, and Roman will image billions more. The bottleneck has moved from data to inference. My program connects simulations and observations through forward modeling and machine learning, so that differences between data and models reflect physics, not method.
Four research themes
Select a theme to see the science questions, key results, figures, and related papers.
THEME 01The Early Universe
The first galaxies and their star clusters seen with JWST: ISM conditions at z > 10 and parsec-scale clumps in strongly lensed galaxies.
THEME 02Structural Evolution of Galaxies
How galaxies build bulges and disks and shut down star formation, inside-out, mapped pixel by pixel from the cosmic dawn epoch to today.
THEME 03Forward Modeling of Simulations
GalSyn turns hydrodynamical simulations into realistic telescope data with known ground truth, so observations and theory meet on equal footing.
THEME 04Machine-Learning Inference
GalSyn-SBI: simulation-based inference that recovers global and resolved property maps and star-formation histories in under a second per galaxy.
The “Pixels to Physics” loop
Cosmological simulations (1) are forward-modeled with GalSyn (2) into realistic multiband observations with known ground truth. These mocks train amortized simulation-based inference networks (3). Applied to JWST, Euclid, Rubin, and Roman imaging, the networks produce resolved maps and star-formation histories for 105–107 galaxies (4). These test physical hypotheses (5) and feed constraints back to the next generation of simulations (6).
Mock and real galaxies pass through the same pipeline, so inference biases can be measured where the answer is known.
Looking ahead
Four connected thrusts. Each pairs a falsifiable science goal with a data-science innovation the survey era needs.
When, where, and how do galaxies shut down?
Is the map of quenching routes across mass and cosmic time a single pattern that scales with epoch, or did the early Universe quench galaxies through routes that are absent later?
SBI applied to all public JWST/NIRCam imaging in five deep fields: ~3×105 galaxies at 1.5 < z < 8. Training on GalSyn mocks from TNG, EAGLE+FLARES, SIMBA, and THESAN tests four hypotheses on outshining biases, the origin of the first massive quiescent galaxies, transition masses, and temporary low-mass quenching.
From JWST to the survey era: a resolved census with Euclid, Roman, and Rubin/LSST
How do environment and the cosmic web shape where star formation dies inside galaxies, and how do bulges, disks, and faint outskirts assemble?
The JWST archive serves as an observed truth set. Degrading it to Euclid and Roman resolution calibrates survey-trained networks before they are applied to 106–107 galaxies. The data-science work covers sim-to-real domain shift, ML deblending, photo-z marginalization, and spatially varying PSFs.
Using resolved galaxies to constrain galaxy-formation physics
Can resolved galaxy populations discriminate between, and quantitatively constrain, the sub-grid physics in simulations?
The plan has three parts: distribution-level comparison of kinetic, thermal, and jet AGN feedback models; SBI over CAMELS feedback parameters using resolved summary statistics; and a JAX-accelerated, differentiable dust module in GalSyn that constrains attenuation curves at high redshift. All of this requires robust inference under model misspecification.
Discovery at small scales and rare phases
Do giant star-forming clumps survive, migrate inward, and build bulges?
A statistical sample of lensed clumps and star clusters from the VENUS and Cosmic Spring collaborations, compared with zoom simulations forward-modeled with GalSyn. ML anomaly detection in SBI latent space will look for short-lived phases such as compaction caught in the act, rejuvenation, and the hosts of compact red AGN.
Next milestones
- Outshining & early quiescent galaxiesGalSyn multi-simulation library.
- Transition masses & low-mass quenchingPublic JWST resolved-map catalog; Euclid/Roman calibration set; ML deblending.
- Environment & resolved quenching with Euclid and RomanHigh-z dust-law inference with differentiable forward modeling.
- Roman resolved inferenceSimulation model comparison; first CAMELS feedback constraints.
- A unified view over 0 < z < 8Release of the full inference stack and GalSyn DR2.
Community infrastructure
// planned public releases
- GalSyn DR2: a benchmark library of synthetic HST, JWST, Euclid, Rubin, and Roman observations from multiple simulation families, with ground-truth maps, SFHs, and quenching labels.
- GalSyn-SBI: pre-trained, calibrated SBI models released alongside GalSyn and piXedfit, with simulation-based calibration, coverage tests, and out-of-distribution flags.
- Foundation-model connection: testing whether multimodal embeddings (e.g., AstroCLIP) fine-tuned with SBI heads improve robustness to domain shift.