2022 Simons Collaboration on Learning the Universe Annual Meeting

Greg Bryan
Columbia University

The Collaboration in Year One

Collaboration Director Greg Bryan will review the status of collaboration efforts, with a particular focus on near-term (and long-term) deliverables and areas that need additional efforts. He will also lead a discussion on collaboration structure and infrastructure. Finally, Greg will provide updates from the Black Hole working group, briefly reviewing the various current projects and discussing where we hope to be in year two.

Shy Genel
Flatiron Institute

Status of the Training Set Generation Working Group and CAMELS

Shy Genel will review the current status of the TSG efforts and CAMELS on a number of fronts. This includes a review of the wide range of simulations already carried out (many in addition to the original CAMELS data release), as well as a discussion of possible future training set runs. Finally, there will be an update and discussion on a recent joint LtU/CAMELS effort to develop novel ways to intelligently choose the parameter space coverage (in both cosmological and astrophysical parameters) for the expensive simulations that maximize our ability to train emulators or constrain initial conditions/parameters (without “wasting” a lot of simulations on completely unrealistic universes).

Jens Jasche
Stockholm University
and
Guilhem Lavaux
Institut d’Astrophysique de Paris

Progress of the BORG Working Group: Success and Challenges for 2023

The BORG method and software have already proven some success in the analysis of galaxy surveys such as 2M++ and SDSS. The working group is now seeking to make the inference sufficiently robust, fast and accurate at intermediate scales (1-10 Mpc/h) to allow for sampling cosmological parameters jointly with the initial conditions. We will highlight recent progress on that front and the promises of the work plan for the coming year. These activities are notably synergistic with the TSG working group, the Accelerated Simulation working group and the Implicit Likelihood Inference working group.

Eve Ostriker
Princeton University

Interstellar Medium, Star Formation, Galactic Winds: Resolved Simulations & Subgrid Models

The evolution of galaxies, including the stellar component that is the main tracer of matter in the universe, depends on physical processes that occur within the interstellar medium (ISM). Gravitational collapse of ISM gas leads to star formation (SF), and then energy is returned to the ISM from recently formed massive stars. Since the return of energy sets the ISM turbulent, thermal and magnetic pressure, this regulates future star formation. At the same time, energy return (in the form of stellar winds and radiation, supernova explosions and cosmic rays) drives outflows of warm and hot thermal gas and relativistic particles into the circumgalactic medium. The mass and energy flows out of galaxies as galactic winds (GWs) reduce future star formation over long timescales. While these processes are known to be essential to the evolution of galaxies, it is not possible to include them directly in cosmological simulations of galaxy formation due to limited spatial/mass resolution. To date, cosmological simulations as well as semi-analytic models have primarily relied on subgrid models for SF, the ISM and GWs with simple functional forms and parameters that are set via empirical tuning. This means that the galaxy formation simulations and semi-anlytic models are not fundamentally predictive; they may obtain the “right answer” for the wrong reason, and/or they may mask issues with standard cosmological theory. A key goal of the LtU collaboration is to move from current subgrid models of ISM, SF and GW processes that are empirically tuned to new models that are based on calibrations from resolved radiation-magnetohydrodynamic simulations that directly treats the relevant physics. The simulations we are using include star-forming ISM “patch” models with a range of galactic conditions, as well as global simulations in dwarfs via cosmological zooms to reach comparable resolution (~1–10 pc) and global simulations in spirals that have a resolution of 10–100 pc. The first two types of simulations enable us to directly model the required physics, with “patch” simulations providing full exploration of parameter space and global dwarfs testing sensitivity to geometry and assessing wind propagation on scales beyond a few kpc. The third type of simulation provides a bridge to lower-resolution cosmological galaxy formation simulations. In addition to the simulations, we are working on algorithms that use robust cosmological variables (i.e., variables that are well resolved in cosmological models) as inputs to calibrated SF rate predictors, and on algorithms for launching and recouping multiphase winds. Eve Ostriker will provide an update on the work to date that has been accomplished by the working group members on the above tasks, and on plans for the coming year. Ostriker will also discuss networking with the other WG to implement our models.

View Slides (PowerPoint)

Rachel Somerville
Flatiron Institute

Status of the Observational

Rachel Somerville will discuss current plans for the generation of a set of pipelines (developed in conjunction with many of the other working groups) to create mock galaxy and secondary CMB maps. These include a stellar synthesis pipeline (SynthObs1) which takes snapshots from cosmological hydrodynamic simulations and generates dust-free (and dust-extinct) stellar and nebular emission properties in specific bands. A second pipeline (SynthObs2) will use the first pipeline to create light cones and realistic photometric data. A pair of addition pipelines (SynthObs3 and SynthObs4) will generate CMB secondary anisotropies from cosmological hydrodynamics simulations and then project to the sky with realistic observational effects.

View Slides (PDF)

Ben Wandelt
Institut d’Astrophysique de Paris

Learning the Universe with Implicit Inference

Implicit inference is an approach to doing Bayesian statistics that has unlocked a large class of previously intractable problems. Implicit inference can work even when the likelihood and/or the prior are intractable distributions because it only requires the ability to generate parameters and data of interest. It is made possible through multiple recent advances in machine learning and deep learning: the efficient representation of multivariate probability density functions; fast generative models for parameters, signal and data; and a dictionary that allows us to translate posterior inference into optimization problems that can be solved using stochastic gradient descent. Ben Wandelt will outline the plan to employ these techniques to “Learn the Universe” and identify outstanding problems as well as promising steps towards their solutions, such as: the generalization to very high-dimensional parameters space; integrating physics priors and constraints with machine-learning approaches; robustness to model imperfections; and validation of the resulting inferences.

View Slides (PDF)