Research
Current research at Princeton
At Princeton, I study how galaxies form, evolve, and enrich the universe around them. I am especially interested in coupling detailed cosmological hydrodynamical simulations with fast, physically motivated semi-analytic models. Hydrodynamical simulations let me follow the complex motion, heating, cooling, and chemical evolution of gas, while semi-analytic models make it possible to explore much larger volumes and a wider range of physical assumptions. By bringing these approaches together, I aim to build models that are both physically informative and computationally efficient. High-performance computing is an important part of that bridge: I run, analyze, and optimize large simulations while developing faster models for broad parameter exploration.
Many of my science questions center on galaxy metal enrichment and the interplay between galaxies and their local environments. Stars create heavy elements, but feedback from supernovae and galactic winds can carry those metals far beyond a galaxy’s visible boundaries. I want to understand where those metals go, how they move through the cosmic web, and how environment influences a galaxy’s ability to accrete gas, form stars, and drive outflows. To me, galaxy formation is a two-way conversation: galaxies continually reshape their surroundings, while those surroundings help determine how the galaxies grow.
I am also fascinated by the epoch of reionization, particularly its temporal and spatial patchiness. Reionization did not happen everywhere at once; different regions encountered ionizing radiation at different times and experienced different ultraviolet intensities and thermal histories. I investigate how this uneven process affected gas accretion and retention in low-mass halos, and how those early local conditions may still be reflected in the dwarf-galaxy population we observe today. Together, these interests connect small-scale baryonic physics to the large-scale structure and history of the universe.
GPU acceleration of RAMSES
At the NVIDIA Open Hackathon at Princeton University in June 2026, I joined collaborators from Princeton, Stanford, and NVIDIA to accelerate RAMSES, an adaptive-mesh-refinement code that combines Godunov hydrodynamics, multigrid gravity, particle evolution, and dynamic mesh refinement. The goal was straightforward but ambitious: move the most computationally expensive parts of a full cosmological calculation efficiently onto NVIDIA GPUs and make the different solvers work together, rather than optimizing each component only in isolation.
We used NVIDIA Nsight Systems and Nsight Compute (ncu) to trace execution, measure GPU utilization, and locate performance bottlenecks. For the hydrodynamics solver, we compared the cube and paper kernel designs, experimented with larger computational stencils, and studied the balance between shared-memory use and throughput on A100 and H200 GPUs in both single and double precision. For the multigrid gravity solver, we fused kernels, reduced device-to-host and host-to-device memory transfers, and minimized accesses to main GPU memory. The resulting multigrid V-cycle reached roughly 2 billion cell updates per GPU per second on an A100 and 4 billion on an H200 in single precision.
We also profiled the particle pipeline, including radix sorting, cloud-in-cell mass deposition, the Poisson solve, and particle position and velocity updates. This exposed scaling limits in the sort and contention in mass deposition, motivating CUB-based sorting and a new cloud-in-cell deposition scheme. By the end of the hackathon, the team brought isothermal hydrodynamics, dark matter, and adaptive mesh refinement together in an integrated GPU-accelerated run.
The CAMELS-RAMSES model
The CAMELS project - Cosmology and Astrophysics with MachinE Learning Simulations - is a large public laboratory of virtual universes. Instead of simulating only one best-guess universe, CAMELS systematically varies cosmological parameters, galaxy-formation physics, feedback strengths, and initial conditions across thousands of simulations. Hydrodynamical runs are paired with gravity-only counterparts, making it possible to isolate the effects of baryonic physics. The resulting galaxy catalogs, summary statistics, and maps of dark matter, gas, and stars provide a rich training set for machine-learning models and emulators that connect physical parameters to observable structure.
I contribute to the development and validation of CAMELS-RAMSES, which introduces the adaptive-mesh-refinement code RAMSES and its galaxy-formation model into this broader simulation ecosystem. RAMSES offers a numerically independent complement to the moving-mesh and particle-based methods already represented in CAMELS. Comparing the same physical questions across independent codes helps us identify which predictions are genuinely robust and which may depend on a particular numerical method or subgrid prescription.
Our work benchmarks the cosmic star-formation history, galaxy and halo mass functions, stellar-to-halo mass relations, and matter maps against IllustrisTNG, Simba, Astrid, and Swift-EAGLE. We then test how these predictions respond to changes in cosmology, stellar and active galactic nucleus (AGN) feedback, and star-formation parameters. The broader goal is to understand which observables carry the clearest information about cosmology and galaxy physics, quantify code-to-code systematics, and build reliable simulation-based tools for parameter inference.
Local filtering mass during inhomogeneous reionization
Reionization heated the diffuse gas between galaxies, increased its pressure support, and made it harder for low-mass dark-matter halos to accrete and retain the fuel needed to form stars. This suppression is often summarized by a single, redshift-dependent filtering mass. Real reionization, however, was patchy: dwarf-galaxy progenitors encountered ionization fronts at different times and experienced different ultraviolet intensities, spectral shapes, source geometries, thermal histories, and filamentary gas supplies. My project asks a more local question: at fixed halo mass and redshift, which parts of a dwarf progenitor’s radiation and thermal history best predict whether it keeps its baryons and continues to accrete gas? Rather than assuming one universal threshold, I am exploring how the characteristic suppression scale varies from region to region and retains a memory of local reionization.
I am developing a set of controlled cosmological zoom experiments with RAMSES-RT, which follows photon transport, non-equilibrium ionization, and gas dynamics together. The initial simulations use a representative dwarf-galaxy progenitor as a test bed for refining the numerical setup and identifying the most informative comparisons. The broader plan is to compare hydrodynamical control runs with radiation-hydrodynamic variants that change the arrival time and character of the ionizing radiation. At this stage, the emphasis is on building a robust experimental framework rather than drawing conclusions from any single target halo.
The analysis will connect the evolving local radiation field to baryon retention, cold gas and inflow, self-shielding, and the thermal history of the growing dwarf. These measurements are intended to help distinguish between gas being photoevaporated from the halo, circumgalactic gas being prevented from cooling, and the loss of fresh filamentary accretion. Ultimately, I want to determine whether a locally informed description of reionization can explain some of the diversity among low-mass halos that is hidden by global, redshift-only models. This project remains an active work in progress, and both the simulation design and the range of targets will continue to evolve as the first experiments guide the next steps.
On-the-fly mapping of intergalactic metal enrichment
Galaxies are cosmic metal factories, but they do not keep everything they make. Supernovae and stellar feedback launch enriched winds that carry heavy elements beyond the interstellar and circumgalactic media and into the intergalactic medium (IGM). Many fast semi-analytic models simplify this cycle by treating newly accreted gas as pristine or assigning it the same fixed metallicity floor everywhere. That approximation erases the patchy, time-dependent nature of enrichment: gas near an active group of galaxies should have a very different chemical history from gas in a quiet cosmic void. To restore that environmental memory without paying the cost of a full hydrodynamical simulation, I developed an on-the-fly enrichment framework that couples RAMSES dark-matter-only halo catalogs and merger trees to the GRUMPY semi-analytic galaxy-formation model.
The framework evolves gas mass and metal mass on a shared 2563 Eulerian grid alongside the galaxy population. At every timestep, GRUMPY calculates how many metals a galaxy produces, how much enriched gas is expelled, and how far a momentum-driven wind can carry it. The ejected material is deposited with normalized, physically motivated kernels, including a Fermi-Dirac diffusion edge that creates a smooth transition between enriched bubbles and the surrounding low-metallicity IGM. Each halo carries a persistent bubble that follows its trajectory through the cosmic web: when the halo moves, the model removes its previous grid contribution and redeposits the complete bubble at the new position. This procedure conserves metal mass exactly, prevents artificial trails along halo orbits, and allows bubbles from neighboring galaxies to overlap naturally. A fast spatial search then measures the mass-weighted IGM metallicity within each halo’s virial radius and feeds it back into the metallicity of newly accreted gas, closing the loop between galaxy growth, outflows, environmental enrichment, and later reaccretion.
The resulting maps recover the broad geography of cosmic enrichment - metal-rich filaments and nodes surrounded by underdense, weakly enriched voids - in qualitative agreement with a CAMELS hydrodynamical simulation using the same initial conditions, but at a fraction of the computational cost. Halo environments become more enriched and chemically diverse with cosmic time, demonstrating hierarchical pre-enrichment as bubbles grow and overlap. For classical dwarf galaxies with stellar masses of roughly 106-1010 solar masses, however, evolving the IGM metallicity has little effect on the stellar or gas-phase mass-metallicity relations because internal metal production remains dominant. Wind experiments reveal the expected trade-off: stronger mass loading removes more metals from galaxies and enriches a larger IGM volume. Mock Lyman-alpha absorbers generated from 10,000 random sightlines reproduce the overall trend between metallicity and hydrogen column density but miss the near-solar tail, which is dominated by dense inner circumgalactic gas that an IGM-only grid intentionally does not model. Finally, an independent analytical reconstruction reproduces the numerical metallicity distribution and explains two characteristic peaks: one near log10(Z/Zsun) = -2.5, driven mainly by the enriched volume surrounding halos of about 1012 solar masses, and a second near -1, where many halo masses converge on the maximum bubble metallicity and form a mathematical caustic. The paper is in preparation for submission to MNRAS.
Galaxy formation and galactic outflows in cosmic voids
Cosmic voids occupy most of the universe’s volume, yet they contain very little matter and far fewer galaxies than filaments, groups, or clusters. Their relative isolation makes them unusually clean laboratories for asking how environment changes galaxy formation - and how effectively the few galaxies living there can pollute an otherwise nearly pristine region with metal-rich outflows. Working with Romain Teyssier, Ulrich P. Steinwandel, and Alice Pisani, I used RAMSES hydrodynamical simulations to ask a simple question with a surprisingly nontrivial answer: how empty does a cosmic void remain once its galaxies begin forming stars and launching winds?
We began with a 1 cGpc/h dark-matter simulation and searched 1,000 randomly placed spheres for an exceptionally underdense target. The selected void has a density contrast of approximately -0.8 and an effective diameter of about 120 cMpc/h. I then simulated its central 25 cMpc/h region at multiple resolutions, expanded the zoom volume to 50 cMpc/h to test whether galaxies near the larger void boundary influence the center, and ran a companion 25 cMpc/h field simulation with mean cosmic density. These controlled comparisons separate three effects that are often tangled together: numerical resolution, the size of the simulated region, and the physical influence of environment. The central void-galaxy population remained stable when the zoom volume was enlarged, showing that the inner 25 cMpc/h region was well isolated, while higher resolution substantially improved the recovered abundance and stellar masses of low-mass galaxies.
Because even an expensive hydrodynamical simulation cannot resolve every faint galaxy, I built an analytical framework that combines the environment-dependent halo mass function, the fraction of halos able to host luminous galaxies, the stellar-to-halo mass relation, and a momentum-conserving supernova-wind model. Calibrating this model to the RAMSES simulations lets us extrapolate down to the filtering mass of star-forming halos and estimate the galactic-wind volume filling fraction - the fraction of space reached by wind-blown material - at effectively infinite resolution. We predict that winds fill about 18.6% of the field volume but only 3.1% of this extreme void. Although dwarf galaxies can drive strong winds individually, their small enrichment volumes and the scarcity of galaxies in voids make their combined contribution surprisingly modest; resolving halos down to about 1010 solar masses is sufficient for a robust total. The prediction also depends strongly on void underdensity and cosmic variance, while void size matters less except for small regions. Applying the same framework to the Local Group void, with a density contrast near -0.5 and a radius of 20 cMpc, gives a filling fraction of 9.6% ± 3.3%. This work was published in MNRAS in June 2025: Much Ado About Nothing: Galaxy Formation and Galactic Outflows in Cosmic Voids (arXiv).
Satellite galaxies in the Merian Survey
Satellite galaxies are small, but they carry an outsized amount of information about how their host galaxies assembled, how dark-matter halos are populated, and how environment shuts down star formation. Our own Milky Way has two unusually prominent, star-forming companions - the Large and Small Magellanic Clouds - but it has long been difficult to tell whether this configuration is common or peculiar. Working with Jenny Greene, Shany Danieli, and the Merian collaboration, I led a statistical census designed to place the Milky Way in a much larger and more representative family of comparable systems.
The Merian Survey is particularly well suited to this problem. It combines two custom DECam medium-band filters, N708 and N540, with deep Hyper Suprime-Cam grizy imaging over a wide field. The filters catch H-alpha and [O III] emission from galaxies at redshifts of roughly 0.06-0.10, sharpening photometric-redshift precision to about 0.01 and allowing us to identify bright, star-forming dwarfs without requiring complete spectroscopy. For the satellite analysis, I selected 393 central Milky Way analogs with host stellar masses of 1010.5-1010.9 solar masses at redshifts 0.07-0.09. I searched from 0.1 virial radii to the edge of each host halo and combined magnitude, color, angular-size, photometric-redshift, and size-mass cuts with masking, visual deblending, and statistical background fields. This produced 793 candidates, or 451 ± 47 star-forming satellites after background subtraction, down to an approximate stellar-mass limit of 108 solar masses. Because the selection favors emission-line galaxies and crowding hides some objects within about 30 kpc of a host, the analysis also tracks incompleteness from quenched satellites and the bright central regions of the hosts.
The census shows a wonderfully broad range of satellite systems: 51 ± 5% of the Milky Way analogs have no detected massive star-forming satellite, 19 ± 4% have one, 13 ± 4% have two, and 17 ± 4% have three or more. Satellite abundance rises most clearly with the total stellar mass of the system, as expected if more massive halos build richer entourages through hierarchical growth. After completeness correction, the average radial distribution is less centrally concentrated than the Milky Way’s satellite population; a simple power law describes the Merian profile better than the best-fitting NFW model, although the most centrally concentrated half of the Merian systems looks considerably more Milky-Way-like. Together, these results suggest that hosting two bright companions is not especially anomalous, even if the Magellanic Clouds’ particular masses, orbits, and ongoing interaction remain distinctive. This work first appeared on arXiv in December 2025 and was published in The Astrophysical Journal in March 2026: The Merian Survey: A Statistical Census of Bright Satellites of Milky Way Analogs (arXiv).
I also contribute to the broader Merian survey and first data-release analysis, including the calibration of photometric redshifts for emission-line dwarf galaxies. In a related non-parametric study of H-alpha morphology, we developed a continuum-subtraction method for the N708 filter, constructed resolved H-alpha maps, and measured asymmetry, Gini, and M20 statistics for 1,250 galaxies. We found that both stellar and H-alpha asymmetry increase with specific star-formation rate. The lowest-mass, most actively star-forming dwarfs tend to concentrate their H-alpha emission into a small number of compact, uneven clumps, whereas quieter dwarfs have more diffuse emission - a vivid sign that dynamical disturbances and centrally compressed gas may help ignite dwarf-galaxy starbursts. My ongoing Merian work uses the same medium-band leverage to improve photometric-redshift selection and extend studies of star-forming dwarfs across the survey’s large volume.
Earlier research
Before I start grad school, I sampled many fields of astrophysics. I put an emphasis on studying ultra-faint dwarf galaxies (UFDs) - one of the oldest, most dark-matter dominated, and chemically primitive stellar systems in the universe - because I was intrigued by their power to challenge the ΛCDM paradigm. UChicago (where I did my undergrad) happens to have many experts in this field. My exploration into dwarf galaxies was encapsulated in three research projects, each of which attempted to answer a specific question.
How do dwarf galaxies quench?
I worked with Prof. Andrey Kravtsov and Dr. Christine Simpson on using the state-of-the-art cosmological hydrodynamical simulation Auriga to probe how dwarf galaxies quench (i.e., shut down their star formation) in simulations across different mass scales. Our findings indicate that UV ionizing photons from the epoch of reionization primarily quench low-mass dwarf galaxies. In contrast, intermediate-mass dwarf galaxies are mainly quenched by environmental processes such as ram-pressure stripping. High-mass dwarf galaxies largely remain star-forming as their substantial gravitational potential allows them to resist both stripping and early-universe quenching processes. We then juxtaposed our simulation results with several observations (the Satellies Around Galactic Analogs, or SAGA; the Exploration of Local VolumE Satellites survey, or ELVES) and discovered a consistent alignment. For more details, please refer to my paper.
How can we more efficiently identify member stars of UFDs in observations?
Upon finishing the project above, I became curious about how observers manage to deal with ultra-faint and low-mass UFDs during their surveys. What surveys and techniques do we need to study the resolved stellar population of Milky-Way (MW) UFDs? To explore this, I worked with Prof. Alex Drlica-Wagner, Prof. Alexander P. Ji, and Dr. Anirudh Chiti. We used photometry to identify member stars from three of the MW UFDs: Boötes I, Boötes II, and Segue I.
Traditionally, observers use spectroscopy to identify member stars from their radial velocities and metallicities. However, one major issue with spectroscopy is that it primarily focuses on the central region of the system. Since spectroscopy is both extremely time-consuming and costly, observers typically apply it only to targets that are expected to yield significant results. The center is a good place to look because this region has a higher density of member stars. This approach leaves the outer region of UFDs largely unexplored. However, intriguing signatures that provide information about the evolutionary history of UFDs, such as tidal disruption, merger events, and supernova feedback, are all found on the outskirts of UFDs. How to address this issue and access information on the outer region of UFDs? We turned to photometry for help!
We utilized the Dark Energy Survey Camera’s (DECam) u-band to compute the photometric metallicity of each star within the camera’s field of view. This is feasible because the u-band encapsulates the prominent Ca II K metal absorption line, which can consequently be employed to establish a relationship between u-band brightness and metallicity. Following this, we applied an isochrone cut, proper motion cut, and photometric metallicity cut to select member stars of each UFD.
Upon mapping the spatial distribution of these member stars, we found indications that Boötes I may be undergoing tidal disruption by the Milky Way host. Conversely, Boötes II exhibited a more isotropic distribution of member stars, but it requires additional spectroscopy for confirmation. Furthermore, Segue I seemed to also be experiencing tidal disruption. Interestingly, we identified new member stars of Segue I extending to 5-6 half-light radii of the system. Paper forthcoming. Stay tuned for updates!
How does stochastic star formation affect global properties of dwarf galaxies?
For my honors thesis, I worked with Prof. Andrey Kravtsov on improving a semi-analytical galaxy evolution model called GRUMPY (credit to the Grumpy Cat). The specific improvement we made involved the implementation of stochastic star formation within dwarf galaxies in the model. While the model’s generated dwarf galaxies demonstrated properties consistent with observations, one inconsistency stood out: the model did not produce any faint, blue dwarf galaxies that have been observed in surveys such as SAGA and ELVES. We hypothesized that introducing stochastic star formation into the model could lead some dwarf galaxies to intermittently form stars following the epoch of reionization, resulting in a blue color in the present day. This concept of stochastic star formation in dwarf galaxies is empirically driven. The star formation rate in such galaxies is primarily influenced by a few star-forming giant molecular clouds (GMCs), subject to small number statistics. In the original model, star formation was averaged over 2.5 Gyr—around one-fifth of the universe’s age—thus failing to represent the bursty nature of star formation in dwarf galaxies. To rectify this, we applied the Fourier transformation to the star formation history of each model galaxy, introducing power on small timescale variations to represent stochasticity. Contrary to our expectations, the introduction of stochastic star formation did not result in a population of faint, blue dwarf galaxies. However, introducing stochasticity in metallicity did influence some faint dwarf galaxies to appear bluer. This is because lower metallicity is typically associated with a bluer color. In line with our findings, researchers involved with the ELVES survey also discovered that faint, blue dwarf galaxies in their survey demonstrated early-type morphology, indicating that their blue color wasn’t a product of star formation. We’ll be detailing these findings in a forthcoming paper. Stay tuned!
Other research explorations: cosmology
In the summer of 2022, I was selected as one of the DAAD RISE Germany scholars and visited the Ludwig Maximilian University of Munich. I worked with Dr. Stella Seitz, Anik Halder, and Laurence Gong. Our project focused on compressing the Dark Energy Survey’s (DES) Year 1 and Year 3 two-point correlation function (2PCF) data. Identifying a method to efficiently compress this vast dataset without losing critical information can enhance the way we store and manage the data for further analysis. This approach has the potential to be extended to other large datasets in the future. Stay tuned for more information!
Non-astro research exploration: machine learning!
In the summer of 2023, I was selected as one of the Lambda scholars to work at Mila, the Quebec AI Institute. Currently, I’m collaborating with Prof. Yashar Hezaveh and Prof. Laurence Perreault-Levasseur on an ambitious project. Our goal is to train a neural network capable of transforming low-resolution images into high-resolution versions. This endeavor has substantial implications for astronomy, specifically in the field of strong gravitational lensing. In this domain, different sections of the source object are magnified to varying degrees, making it challenging to accurately represent the source image using a uniform resolution.