Flatiron Institute Center for Computational Astrophysics Pre-Doctoral Program

Matteo Cantiello (Research Scientist)

I am a theorist and in my research I use a variety of computational and observational tools, including 1D stellar evolution (MESA), 3D magneto-hydrodynamics calculations of stellar interiors (Dedalus, Athena++), the study and observation of wave propagation inside stars (Asteroseismology), the observations of stellar populations, stellar explosions, and the gravitational wave emission from the mergers of compact stellar remnants.

Stars in AGN disks
Stars are likely formed in, or captured by, the disks of active galactic nuclei (AGN). The disk conditions profoundly change the star’s evolution, with AGN stars accreting large amounts of mass and becoming massive/very massive. This project could involve either modeling the accretion stream with radiation hydrodynamics software instruments like Athena++, modeling the long-term stellar evolution in the MESA software instrument, or studying the interplay of stellar dynamics, AGN disk models, and evolution, tying together output from a variety of tools with semi-analytic models. Another possible project could involve calculating the rate of visible explosive transients in AGN disks from a population of massive AGN stars. These predictions could be useful for VRO/LSST, LIGO/VIRGO, etc.

Modeling stellar and planetary ingestions with MESA
Planets and stars can be engulfed when, e.g., their host or companion star ascends the giant branch. Using a 1D code (MESA) it is possible to account for the energy deposited during the spiral in and determine the evolution of the primary star. This project aims at creating a grid of light-curves to be used as templates for observations with VRO/LSST.

Julianne Dalcanton (Senior Research Scientist & Director)

I work on using large, multiwavelength datasets of nearby galaxies to constrain the underlying astrophysics of stars and gas in galaxies. These include large HST and JWST imaging programs in Local Volume galaxies, including the Panchromatic Hubble Andromeda Treasury (PHAT) and the Triangulum Extended Region (PHATTER), imaging M31 & M33, a new HST survey of hundreds of peculiar & interacting Arp galaxies, and a new VLA X-Large high-resolution survey of all star forming galaxies in the Local Group.

My interests in working on these projects are very broad, and I am happy to collaborate with applicants on developing project ideas of mutual interest.

Chris Hayward (Research Scientist)

I work on a variety of topics in galaxy formation. I mainly use hydrodynamical simulations, typically those from the Feedback in Realistic Environments (FIRE) project, of which I am a senior member. I also employ analytic toy models, semi-analytic and semi-empirical models, and occasionally observations (usually via collaborations, but I’ve led some observational work myself). Much of my work involves performing radiative transfer to generate synthetic observations to more directly compare simulations and observations (e.g., forward-modeling).

Specific current interests include: star formation self-regulation via stellar feedback and the relation to bursty star formation and outflows; infrared-selected galaxies; dust; and galaxy protoclusters. Previous pre-docs whom I co-mentored have worked on the effect of cosmic rays on thermal instability, developing a neural network-based emulator for radiative transfer, hierarchical modeling of galaxies’ dust attenuation curves, and how well infrared-luminous galaxies trace protoclusters. I’m open to discussing possible projects related to any of the above interests, but especially the following topics:

• dusty star-forming galaxies (DSFGs; with Flatiron Research Fellow Rachel Cochrane);

• JWST synthetic observations (with Flatiron Research Fellows Matt Orr and Rachel Cochrane); and

• stellar feedback-driven outflows: comparing different simulations and observational constraints (with group leader Rachel Somerville).

Shirley Ho (Group Leader)

Research interests: Our group develops and applies state-of-the-art deep learning techniques to solve astrophysical challenges that are not solved by current methods.

Using Transformers (think chatGPT) to accelerate simulations

Turbulence is notoriously difficult to model due to its multi-scale nature and sensitivity to small perturbations. Classical solvers of turbulence simulation generally operate on finer grids and are computationally inefficient. We have built previously Turbulence Neural Transformer (TNT), which is a learned simulator based on the transformer architecture, to predict turbulent dynamics on coarsened grids, and we show that TNT outperforms the state-of-the-art U-net simulator on several metrics. We like to push forward with TNT to multiple conditions in turbulence simulations, and check if it outperforms existing solvers when we generalize to additional simulation datasets.

Create neural network architecture with physical constraint

In scientific disciplines, it is very often we have physical prior and constraints that we know our data needs to satisfy. However, very often these constraints are not obvious in the real-world data as we only observe parts of the sky, or part of the data is missing. In this project, we would like to design neural networks that include physical constraints that we know exist, while allowing for the data to be imperfect. This will follow up on our work on Lagrangian Neural Network.

Yan-Fei Jiang (Associate Research Scientist)

I am generally interested in numerical simulations of accretion disks around different kinds of systems, including compact objects and planets, particularly related to the effects of thermodynamics on disk structures. Examples include in structures of AGNs, migration in protoplanetary disks, effects of dust on disk dynamics. I am also interested in stellar structures, particularly massive stars, cosmic ray and radiation feedback in galaxies, as well as numerical simulations of transient systems such as TDEs.

Chirag Modi (Flatiron Research Fellow), Shirley Ho (Group Leader) and Ben Wandelt (IAP/CCA)

I am broadly interested in astrophysics and statistical inference. I primarily focus on developing forward modeling approaches for cosmological analysis, including Bayesian hierarchical modeling and simulations-based inference. I am also generally interested in exploring well-motivated problems in astrophysics that can benefit from using machine learning. I also have a joint appointment with the Center for Computational Mathematics (CCM) where I think about new MCMC and variational inference algorithms.

Bayesian hierarchical modeling (or field level inference) for cosmological analysis
Broadly, the goal here is to model and use the full field-level cosmological data for analysis, instead of using any summary statistics. To be able to do this, we need to use differentiable simulations and simultaneously infer initial conditions and cosmology parameters. A lot of work has been done, and remains to be done in this, but some immediate projects might look as follows:

• developing new normalizing flow architectures for cosmology data, combining them with Bayesian inference algorithms and do first joint inference of parameters with the initial conditions;
• developing a pipeline to include survey systematics like fiber collisions and survey masks in modeling the field level data; and
• combining different cosmological observables like galaxy clustering and weak lensing to do joint inference.

The final project will be tailored in discussion with, and based on, the interests of the student.

Simulation-based inference for cosmological analysis
Simulation-based inference has emerged as a powerful technique to use new summary statistics, which lack analytic models and likelihood, for cosmological analysis. However, since cosmological simulations are not 100 percent accurate, developing a ‘robust’ simulation-based inference (SBI) framework in cosmology is still an open question. Some of the questions that can be tackled in this project are as follows:

• How can robust machine learning algorithms account for model misspecification i.e. when the training data and observed data are not generated from the same simulation?
• How do we detect that our model is misspecified and hence the inference is not trustworthy?
• How do we develop ways to combine SBI with traditional analysis, like power spectrum analysis on large scales. Does it make analysis more robust?

The final project will be tailored in discussion with, and based on, the interests of the student.

Lucia Perez (Flatiron Research Fellow) and Shy Genel (Research Scientist)

I am interested in a variety of areas. Namely: how we attach galaxies to dark matter halos in simulations; how galaxy properties and clustering clues us into how dark matter, cosmology and astrophysics works; how to create big simulations across wide parameter spaces better and faster; and getting better at machine learning and computational tools for inference. Voids, the epoch of reionization and Lyman-Alpha Emitters/narrowband galaxies are additional areas of study.

An unexplored landscape for studying the galaxy-halo connection with machine learning
CAMELS-SAM is a new and large-volume ‘hump’ of the CAMELS (Cosmology and Astrophysics with MachinE Learning Simulations) project that offers many exciting projects with machine learning and better understanding how the galaxy-halo connection varies with cosmology and astrophysical models. The suite includes more than one thousand dark-matter only simulations of (100 h^-1 cMpc)^3 with different cosmological parameters (Omega_m and sigma8) and run through the Santa Cruz semi-analytic model for galaxy formation over a broad range of astrophysical parameters. The larger volume has already enabled studies using galaxy clustering in Perez, Genel, et al. 2022, and others outside of the CAMELS collaboration. Opportunities exist to create new data sets with the CAMELS-SAM infrastructure, such as incorporating other SAMs or methods of populating dark matter halos with galaxies or applying post-processing pipelines to generate observables. The suite is ripe for machine learning studies, such as using convolutional or graph neural networks to utilize large-volume field-level information and testing other summaries or tools for cosmological parameter inference.

Potential applicants are encouraged to reach out to Perez and/or Genel to discuss possible projects.

Adrian Price-Whelan (Associate Research Scientist)

The CCA Galactic Dynamics group is open to applications for predoctoral program projects to work with Adrian Price-Whelan and Flatiron Research Fellows Nico Garavito-Camargo and Danny Horta, especially on projects that overlap with the broad interests of the group:

• Searching for signatures of dark matter substructure in the Milky Way (through, e.g., modeling the density structure of stellar streams),

• Quantifying the impact of time-dependent phenomena on the dynamical evolution of galaxies using numerical simulation and new analysis methods,

• Studying the global distortion of dark matter around the Milky Way due to the infall of the Magellanic Clouds and other satellite galaxies,

• Connecting stellar element abundance data with kinematic measurements to improve dynamical inferences,

• Interpreting kinematic signatures of resonances and time-dependence in the Milky Way using Gaia and SDSS-V data,

• Developing new methods to analyze stellar streams around external galaxies in anticipation of deep imaging data from the Rubin Observatory/LSST and the Roman Space Telescope.

Asymmetric mass loss from star clusters (with Sarah Pearson (NYU) and Mordecai-Mark Mac Low (AMNH))
Why do some star clusters in the Milky Way form stellar streams while others don’t? With the Nbody code PeTar, designed to model collisional stellar systems, we propose to test whether interactions between globular clusters (GCs) and giant molecular clouds (GMCs) in the disk of the Galaxy can lead to asymmetric shocks and preferentially cause some GCs to disrupt while others of similar masses and on orbits do not.

Douglas Rennehan (Flatiron Research Fellow) and Chris Hayward (Research Scientist)

I’m generally interested in the evolution of gas within hot halos at all redshifts, but I currently focus on the high redshift Universe. Specifically, I am interested in the formation of the intracluster medium and its evolution in galaxy clusters up to z = 2. Numerically I have been involved in creating new black hole feedback models for cosmological simulations, as well as sub-grid turbulence models.

Ram pressure stripping in the early universe (with Stephanie Tonnesen, Research Scientist)
Satellite galaxies that move through the hot gas within galaxy clusters — the intracluster medium (ICM) — lose hot and cold gas due to a ram pressure, leading to a shut down in star formation. How these systems quench and evolve is a deep topic that has received much attention. However, most of the work has been done on systems that are in equilibrium — virialized galaxy clusters in the low redshift universe. In the low redshift systems, the gas is in hydrostatic equilibrium while the satellite galaxies are orbiting as effectively tracers of the underlying potential. In higher redshift galaxy clusters, it is not clear that the gas is in equilibrium, the gas densities are much higher, and the satellite galaxies may be on their first infall or moving with the bulk flow. For this reason, ram pressure may act on different timescales depending on: (a) the bulk velocity of the gas in the cluster, and (b) the motion of the galaxies through the medium.

We want to investigate how ram pressure stripping rates change in the early evolution of galaxy clusters in order to understand satellite galaxy quenching in these systems.

The project would involve analyzing a set of pre-run zoom-in simulations of galaxy cluster simulations to z = 2, run with the GIZMO hydrodynamics+gravity code. All of the analysis can be done in Python.

Rachel Somerville (Group Leader) and Chirag Modi (Flatiron Research Fellow)

We are offering jointly supervised projects at the intersection of cosmology, galaxy formation and machine learning.

One of the major obstacles to robust cosmological inference is the difficulty of accurately modeling galaxy formation in very large volumes, for comprehensive suites of cosmological parameters. We are offering projects that will contribute to the development of computationally efficient methods to create physics-motivated predictions for observable galaxy properties, as well as to disentangling the roles of astrophysics and cosmology on galaxy formation. Some examples of possible projects include:

• Develop machine learning based methods to rapidly create merger trees for large volume cosmological simulations. We have already developed a method based on recurrent neural networks, which has been shown to work well in a Planck-like LCDM cosmology. This method could be generalized and tested across a wider range of cosmologies, and implementation within ‘cheap’ lower resolution simulations such as particle mesh could be developed and tested in combination with semi-analytic models for galaxy formation. Alternatively, we could explore a generative Graph Neural Network implementation of the merger tree building algorithm.
• Learning the DNA of galaxies. Intrinsic and observable galaxy properties (such as stellar mass, star formation rate or luminosity) are determined by an intricately intertwined — and poorly understood — combination of astrophysical processes and the backbone of structure formation governed by cosmological parameters. A possible project area would make use of the CAMELS simulation suite, which has explored a huge range of both astrophysics parameters, different implementations of astrophysics subgrid recipes and cosmological parameters. We could explore questions such as: Can we disentangle the impact of astrophysics from cosmology for various galaxy properties? Which properties of the total matter density/velocity field history are most predictive of galaxy properties?

We are happy to work with interested students to craft a project that best fits their interests and learning goals.

Francisco Villaescusa-Navarro (Research Scientist)

I am interested in a variety of areas. Namely: machine learning, cosmology, astrophysics, numerical simulations and high-performance computing.

ML with cosmo-astro
I’m interested in machine learning projects applied to cosmology and/or astrophysics in the broad term.

Jiayin Dong (FRF), Phil Armitage (Group Leader; CCA/Stony Brook), and Dan Foreman-Mackey (RS)

Understanding early planetary system dynamics using exoplanet populations

We are interested in linking theoretical models of early planetary system evolution to data on giant planet populations (eccentricities, obliquities, orbital radii, etc). Multiple mechanisms — including disk-driven migration, planet-planet scattering, and secular dynamical processes — modify the architecture of planetary systems after the planets form, and can lead to the formation of the hot and warm Jupiter populations. We would like to see if it is possible — using N-body simulations to forward model the dynamics and modern statistical methods to compare against data — to use the large population of known planets to robustly determine which of these processes is most important.

Astroplasma Group

Researchers in our Flatiron Institute group are broadly interested in the plasma physics of neutron stars and black holes. Student projects can be done in collaboration with CCA Flatiron Research Fellows Jordy Davelaar, Joonas Nattila, Libby Tolman and Vladimir Zhdankin. Lorenzo Sironi (CCA/Columbia) will act as a senior mentor for the projects.

Magnetic reconnection in black hole accretion flows
Accretion flows around black holes can contain regions with strong magnetic field lines and highly magnetized plasmas. These regions are prone to impulsive and explosive magnetic reconnection. We will use radiative kinetic simulation to model this phenomenon in the largely explored regime of ultra-fast cooling. The student will work closely with us to perform 2D and 3D simulations of the reconnection and analyze the results. Understanding the plasma dynamics in this new regime is important for interpreting puzzling observations of flares seen in the accreting systems, such as the black hole in the center of our own galaxy.

Using computer vision for automatic segmentation of turbulent plasmas
Computer vision and machine learning tools offer an exciting new way for automatically analyzing and categorizing information from complex computer simulations. The student will work closely with us to apply state-of-the-art computer vision techniques to build an analysis framework to automate the detection of ubiquitous plasma structures like plasmoids and current sheets. We will then apply the framework to simulations of turbulent plasmas around black holes and neutron stars.

Max Isi (Flatiron Research Fellow) and Will Farr (Group Leader; CCA/Stony Brook)

I am a gravitational-wave astrophysicist specializing in the analysis of LIGO-Virgo data. I primarily think about how we can use LIGO-Virgo signals to learn about the nature of black holes, in the context of both astrophysics and fundamental physics. I also think about the signal detection and characterization process more generally, and how we can extract the most information out of our data. Besides binary black hole mergers, I have developed methods for continuous waves expected from pulsars in the galaxy, stochastic backgrounds due to faraway sources, and unmodeled transients that could be produced by supernovae or other unknown sources. I spend a bit of my time thinking about future detectors, like the Cosmic Explorer or LISA.

Measuring the properties of heavy black holes in LIGO-Virgo binaries
The collection of binary black holes detected through gravitational waves by LIGO and Virgo contains an increasing number of heavy, far away systems which are of high interest in understanding the astrophysics of black hole formation and their role in galaxies. Unfortunately, heavy black hole mergers occur at low frequencies, often close to the edge of our detection sensitivity, causing these signals to be short and difficult to analyze. Nonetheless a number of puzzling measurements, like the highly precessing spins inferred for the GW190521 binary, seem to indicate that there is more information hiding in these short signals than we might expect. The goal of this project will be to further understand the measurement of black hole properties in heavy binaries, especially their spins, and determine whether the final portion of the merger signal indeed carries imprints of the preceding orbital dynamics, or whether our measurements are suffering from systematics (perhaps due to unmodeled physics, like eccentricity). To achieve this, we will use standard and new inference techniques (some developed inhouse at CCA) to analyze real and simulated gravitational-wave data. What we learn will impact future LIGO-Virgo measurements and their interpretation by the broader astrophysics community.

Melissa Ness (Associate Research Scientist; CCA/Columbia) and Megan Bedell (Associate Research Scientist)

We are interested in a variety of areas. Namely: galactic archaeology, stellar spectroscopy, chemical evolution and planetary ecosystems.

Neutron capture variability: a doppelganger study
Stellar abundances are highly correlated, although for populations with fixed supernovae abundances, the neutron capture elements look to capture information about stellar birth sites, as well as stellar evolution (e.g., atomic diffusion, mass transfer from binary systems).

Using GALAH spectra, in this project we will directly examine the variance of the neutron capture absorption features in open clusters compared to a reference sample of field stars. The results will have implications for the information captured in neutron capture abundances, chemical abundance dimensionality, birth versus ‘evolved’ abundances, and stellar populations

Alice Pisani (CCA/Cooper Union)

I am interested in a variety of areas. Namely: cosmology, large-scale structure, cosmic voids, galaxy evolution and machine learning.

Forecasting the constraining power of cosmic voids in the next generation of surveys
Cosmic voids, the large under-dense regions in the distribution of galaxies with sizes from tens to hundreds of Mpc/h are promising laboratories to extract cosmological information. The unique low-density character of voids makes them extremely sensitive to diffuse components such as neutrinos and dark energy; also, voids represent ideal environments to test modifications of gravity. Upcoming surveys will provide hundreds of thousands of voids (including Roman, SPHEREx, Euclid, DESI, PFS), enabling a comprehensive use of cosmic void statistics (see Pisani et al 2015, 2019). Statistics such as the void-galaxy cross-correlation function and the void size function have been measured from data, their use already provides competitive constraints on cosmological parameters (see e.g. Hamaus et al. 2020, Contarini et al. 2022). I am interested in exploring constraining capabilities from voids from upcoming spectroscopic surveys, but I am also open to collaborate on topics within the broader area of cosmic void science, including intersections with CMB and weak lensing. While a set of potential projects is defined upfront, I am also very open to developing a project idea together. Depending on the considered project, the work can be done with one or more CCA co-mentors.

Project ideas include:

• Accurate forecast for constraints from void statistics to be measured by the Nancy Roman Space Telescope NASA observatory (such as the void-galaxy cross-correlation function and the void size function). The project focuses on investigating the behavior of models at relatively smaller scales to enable a gain in constraints exploiting Roman’s dense sampling.

• Forecast for constraints from void statistics to be measured by the SPHEREx mission. This project aims at exploring the science from a sample of voids at (relatively) low redshift.

Exploring the connection between cosmic voids and galaxy properties
I am interested in exploring galaxy properties within voids, and their impact on void statistics for cosmology. Voids contain few, isolated galaxies, and can be seen as privileged environments to investigate galaxy properties. Additionally, void identification brings spurious voids: I am interested in exploring their characterization to increase signal-to-noise for applications relying on void statistics. While a set of potential projects is defined upfront, I am also very open to developing a project idea together. Depending on the considered project, the work can be done with one or more CCA co-mentors.

Project ideas include:

• Characterizing galaxy properties inside cosmic voids, and their impact on void statistics.

• Machine learning characterization of void properties, with a focus on the analysis of spurious voids.