2024 Simons Collaboration on Learning the Universe Annual Meeting

The third annual meeting of the Simons Collaboration on Learning the Universe was held in New York City on September 19–20, 2024, bringing together nearly 100 participants to discuss the collaboration’s results and plans. The evolution of the universe is shaped by its initial conditions and physical laws, but these must be inferred from observations rather than directly measured. This collaboration is using a Bayesian forward modeling approach, where initial conditions are repeatedly sampled, their observational consequences are predicted, and the results are compared to real data to infer the most likely initial states. However, this process is highly challenging due to the complexity and cost of galaxy formation simulations and the vast dimensionality of the initial parameter space.

The event opened with a talk by the director, Greg Bryan, who discussed the progress of the collaboration and our efforts and successes in integrating multiple scientific fields to create a framework for understanding cosmological observations and galaxy formation physics. He highlighted the project’s achievements, challenges and future goals, emphasizing the collaboration’s potential for transformative discoveries.

Eve Ostriker reviewed the development of new subgrid models for simulating star formation, galactic winds and interstellar medium dynamics in galaxy formation models. Subgrid treatments are essential because the spatial and mass resolution of cosmological simulations is too coarse to directly follow star formation and feedback effects, and new developments that make use of radiation-MHD simulations with much finer resolution promise to significantly improve the physical fidelity and predictive power of baryonic modeling. Ostriker discussed the steps involved and results obtained to date from calibration, validation and implementation of subgrid models. Initial results that the LtU team has obtained using the PRFM star formation and Arkenstone wind subgrid models, in isolated galaxy and cosmological simulations, were presented. Ostriker also outlined ongoing testing, improvements and future extensions of the treatments of star formation and galactic outflows, and how members of the SFR+ISM+GW and CM working groups are collaborating for these next steps.

Ben Wandelt explored the impact of the cosmic initial conditions on the universe’s structure, presenting new reconstruction techniques and emphasizing the power of combining diverse datasets. In the “Go Big” project, the team applied state-of-the-art machine learning methods to the SDSS CMASS dataset, showcasing key innovations that enable scaling to present-day and future surveys. The first cosmological constraints from CMASS data were previewed, highlighting the potential for comprehensive, multi-probe cosmological inference at an unprecedented scale. Rachel Somerville discussed the challenges in linking theoretical models with observational data and highlighted the team’s progress on development of synthetic observations to interpret galaxy formation and evolution. She emphasized the importance of these tools for analyzing data from upcoming astronomical surveys and their role in the LtU collaboration. In the final talk of the first day, Lars Hernquist focused on improving the modeling of supermassive black holes in cosmological simulations, covering their seeding, accretion and feedback processes. He presented progress within the collaboration and outlined plans to use new models and data from the James Webb Space Telescope to test these theories.

On the second day, Volker Springel highlighted the progress in cosmological simulations, particularly with the use of large-scale numerical models that incorporate complex physics like star formation and black hole dynamics. He also described the development of new algorithms and computational tools that improve the resolution and efficiency of these simulations, enabling more detailed studies of galaxy formation and cosmic structure. Laurence Perreault-Levasseur outlined the development of machine-learning-powered methods to accelerate cosmological simulations. These new techniques allow more efficient and precise modeling of large-scale cosmic structures, improving the ability to interpret upcoming survey data and enabling advanced inference techniques. Guilhem Lavaux presented the Manticore project, an effort to reconstruct the universe’s initial conditions using galaxy catalogs. These reconstructions aim to resolve cosmological tensions and support future missions like DESI and Euclid, advancing the understanding of cosmic structure formation.

Preceding the annual meeting, the team held a more informal collaborative workshop on Sept 16–18, with some 70 participants carrying out joint working-group meetings, hack sessions and planning discussions. With an abundant supply of break-out rooms and coffee, a great deal of progress was made on many ongoing LtU initiatives. In addition, a number of new collaborative projects were launched, including a clear definition of the new flagship LtU cosmological simulations, incorporating new methodology from all of the working groups.

With new subgrid models, machine learning methods and improved simulations, the team aims to resolve cosmological tensions and gain insights to the universe’s origin, the formation of cosmic structures and the evolution of galaxies. Once our techniques are validated on already published data sets, LtU will have the unique capability to provide physical modeling at the scale of survey data to be released from current surveys. Our analyses will test models of galaxy formation and explore tensions and hints of departures from the standard cosmology.

We remain deeply grateful to the Simons Foundation for supporting this collaboration; none of the methodological breakthroughs we have achieved so far would have been possible by individual groups working in isolation. The next year will see Simons Collaboration making even greater strides in cosmological analysis of real data, enabled by bringing together techniques from across the collaboration and propelled by the connections forged in the past few years. We look forward to sharing the results at our annual meeting in 2025.

Greg Bryan
Columbia University

Learning the Universe: Director’s Overview
View Overview Slides (PDF)
View Wrap-Up Slides (PDF)

The Simons Collaboration on Learning the Universe (LtU) has, over the past three years, made enormous progress towards our goal of creating an effective and robust framework to use observations of the universe to simultaneously infer fundamental cosmological properties and to develop new insights into the physics of galaxy formation. To do this, we have assembled and integrated scientists from four very different fields (and a dozen institutions) into a coherent and interactive collaboration that is accomplishing science none of us could have done alone. Having built most of the connections, theoretical foundations and software tools, we are now moving to apply them to our overarching goal. I will provide an overview of our project plan, briefly highlighting the achievements we have made toward this goal and the places where more work is required. I will describe some of the technical, organizational and sociological challenges that we have faced in this project and some of the solutions that we have discovered (as well as some of the mistakes we have made). Finally, I will describe the enormous potential that continued work in this area promises to deliver and describe the collaboration’s vision for the years ahead.
 

Lars Hernquist
Harvard University

Supermassive Black Holes in Cosmological Simulations
View Slides (PDF)

Cosmological simulations have shown that including supermassive black holes is essential to replicate observed galaxy populations. This is especially true for massive galaxies, whose star formation is halted by black hole feedback. However, simulations have approximated these black holes due to incomplete theories on their origin, growth, and feedback processes. In Learning the Universe, we have undertaken an ambitious program to develop more physical treatments of the dynamics of supermassive black holes, how they accrete gas from their surroundings and grow, and the detailed way energy produced by accretion affects gas in and around galaxies. In this talk, I review progress made to date within the LtU collaboration on our three principal goals related to these aspects of modeling supermassive black holes in cosmological simulations and describe a fourth avenue of investigation that grew out of an external collaboration concerning the “seeding” of black holes in simulations. Going forward, we will exploit the progress to date to formulate a new subgrid model for feedback that will be used in the next generation of cosmological simulations of galaxy formation. We will use the novel approach for black hole seeding to test our models for black hole dynamics and fueling by comparing the outcomes of simulations to data from the James Webb Space Telescope. Given the degeneracy between feedback from stellar evolution and supermassive black holes, this will require close coordination with the star formation group and, ultimately, the cosmological modeling group.
 

Guilhem Lavaux
Institut d’Astrophysique de Paris / Centre National de la Recherche Scientifique

Inferring Initial Conditions of the Universe: Creating the Universe’s Digital Twin
View Slides (PDF)

Our present observations of the universe are a direct consequence of the interaction of two ingredients: the physical laws of nature and the initial configuration of gravitational fluctuations in the primordial universe. The Learning the Universe (LtU) collaboration aims to provide insights into these two ingredients by leveraging large data and a high degree of phenomenological understanding at the small scale of these same data. Constructing physically plausible digital twins of our universe from data is the driving motivation behind the BORG working group of the LtU collaboration. At the heart of this endeavour is the Manticore project, an ambitious effort to reconstruct the initial conditions of our universe within a (6 Gpc)^3 volume at 6 Mpc resolution, utilizing gold-standard galaxy catalogues (e.g., 2M++, SDSS main and LRG, SDSS-BOSS) spanning 0 < z < 0.8. This talk will highlight significant modelling enhancements, stringent new tests and compelling early results. Beyond Manticore, we will also showcase new advances in the modelling of the initial conditions of the Local Group. The resulting data products promise to catalyze advancements in supernovae and gravitational wave cosmology, galactic physics, shear lensing and multi-wavelength cross-correlations. We expect that it will help notably at understanding and reducing cosmological tensions between data sets. Finally, the Manticore framework is designed to integrate forthcoming datasets from missions like DESI, Euclid and Roman, ensuring its relevance and impact throughout the next decade.   Eve Ostriker
Princeton University

Star Formation and Feedback: Developing Robust, Physics-Based Subgrid Models for Cosmological Galaxy Formation
View Slides (PDF)

I will review the accomplishments and future plans for the LtU collaboration’s work on modeling star formation, the interstellar medium and galactic winds. We have developed, tested and deployed new subgrid models that are crucial for galaxy formation modeling in both numerical simulations and semi-analytic models. The stellar content of galaxies—the most direct observable on cosmic scales—depends on (1) the rate of star formation in galactic gas, (2) the rate of gas ejection from galaxies due to radiative, supernova and cosmic ray feedback and (3) the rate of gas accretion from the circumgalactic medium. These processes occur on scales far below the resolution of cosmological simulations, making subgrid models necessary. We use high-resolution numerical radiation magnetohydrodynamic algorithms to directly simulate the star-forming interstellar medium, which also naturally produces galactic outflows. We consider a wide range of conditions (varying gas, star and dark matter content, as well as metallicity) and use these simulations to calibrate new subgrid models for the dependence of star formation and galactic outflow rates on galaxy properties measurable at the much coarser resolution of cosmological simulations. The accuracy and robustness of our subgrid models are tested using different types of numerical simulations and through comparisons with observations where possible. We also design methods to implement multiphase winds in cosmological simulations, requiring novel treatments to follow crucial interactions between thermal phases and maintain energy conservation. We collaborate with the cosmological simulation and synthetic observation groups to deploy the models we have developed. In the past year, we have rolled out several key deliverables, with papers published or submitted. This talk will summarize our results in star formation and wind modeling, leading into work to be described by later speakers. Additionally, I will discuss our plans for the coming year, as we ramp up a new simulation framework to include additional physics features, test new subgrid model approaches for larger cosmological simulations and further explore parameter spaces revealed by recent high-redshift observations.
 

Laurence Perreault-Levasseur
Université de Montréal

Accelerated Forward Modeling and Robustness
View Slides (PDF)

Large-scale upcoming survey data will allow us to shed light on fundamental questions such as the nature of dark energy and dark matter, and the properties of neutrinos, gravity and the early universe. However, the full power of these surveys will only be reached if theoretical modelling achieves the same level of quality and accuracy as their data. Cosmological numerical simulations are key to developing such theoretical models and thus are crucial for the interpretation of these datasets. As we attempt to probe the universe with increasing precision and accuracy on increasingly larger scales, however, the computational complexity of simulations quickly becomes intractable. In this talk, I will review our collaboration’s work on developing the next generation of machine-learning empowered simulation methods to accelerate the simulation of large-scale cosmological volumes. To emulate dark-matter only evolution, we have developed a number of gravity solvers, learning from numerical training sets and enabling the advanced inference techniques employed by the collaboration. This includes an exceptionally precise learned symbolic emulator for the linear power spectrum and for the non-linear power spectrum. It also includes physics-informed machine-learning methods to learn corrected evolution equations that improve accuracy without additional computational expense. Moreover, we have developed halo and galaxy emulators, such as the Charmemulator and PineTree, as well as diffusion-based models to learn to imprint probabilistic galaxies on dark matter density fields. Finally, we have developed a suite of statistical methods to ensure the accuracy of our trained emulators that can assess the quality of the distribution of the produced simulations compared to traditional methods, and even to real data, in extremely high-dimensional spaces.
 

Rachel Somerville
Flatiron Institute

From the Box to the Sky: Creating mock observations to enable robust inference in the observational plane
View Slides (PDF)

One of the paramount challenges that we face in extracting scientific insights from observational data is in making robust connections between first-principles simulations and observables. During this first phase of the Learning the Universe collaboration, we have focused on creating realistic mock galaxy surveys similar to the SDSS survey of nearby galaxies, and mock CMB skies similar to observations from the ACT survey. In this talk, I will describe the end-to-end mock galaxy survey pipeline that has been developed by the LtU team, which allows us to produce realistic SDSS-like “lightcones” including predictions for galaxy photometry. A critical part of this effort has been developing a simple but realistic model for how interstellar dust impacts the light that emanates from galaxies. I will describe radiative transfer simulations that have been used to motivate and ground our dust modeling efforts, as well as the new dust model. This will lead to recent work exploring the interplay between the dust parameters and cosmological and astrophysical parameters using our LtU Simulation Based Inference machinery. I will describe the “LtU Connections” flagship project, in which we are laying the groundwork to perform SBI with physics-grounded simulations in the space of observables. On the CMB side, we have developed multiple techniques to turn the output of fast dark-matter only simulations into thermal and kinematic Sunyaev-Zel’dovich CMB lensing signals, which will also be presented. Finally, I will end with our plans for the next phase in connecting simple theory to the observed world.
 

Volker Springel
Max Planck Institute for Astrophysics

The Promise of Next Generation Hydrodynamic Cosmological Simulations
View Slides (PDF)

Cosmological hydrodynamical simulations are an indispensable and uniquely powerful tool to link fundamental parameters of cosmological theories with small-scale astrophysics, thereby allowing predictions of numerous observables far into the non-linear regime. Within the cosmological modelling group of LtU, we seek to build upon the recent successful calculation of IllustrisTNG by expanding the physical faithfulness of the numerical treatments of star formation and black hole growth, as well as their associated energetic feedback processes, and in addition, enlarging the size and statistical power of the leading cosmological models, as this is required to take full advantage of upcoming new survey data. In my talk, I will present results we recently obtained in this research program, for example in the form of the MillenniumTNG simulations, a suite of new, very large volume simulations that make substantial inroads towards interfacing galaxy formation with precision cosmology. I will also review the methodologies we currently pursue to obtain future multi-physics, multi-scale simulations that realize more reliable and thus more predictive simulations. The road towards such next generation of galaxy formation simulations is rich with challenges and opportunities and is profoundly intertwined with current technological trends, many of them we actively pursue within LtU.
 

Ben Wandelt
Institut d’Astrophysique de Paris

Implicit Likelihood Inference, Go Big, and Beyond
View Slides (PDF)

To achieve the LtU goals of transforming cosmological data analysis requires scaling simulation and analysis techniques beyond the capability of traditional methods. The implicit likelihood inference group has pioneered cutting-edge, bespoke machine learning methods for high-dimensional inference with nonlinear simulations. I will showcase our flagship “Go Big” project, where we are applying these techniques to the SDSS CMASS North Galactic Cap dataset — the largest dataset ever analyzed using simulation-based, implicit likelihood techniques. This analysis exceeds previous proofs of principle by an order of magnitude in data volume. I will detail the key innovations that enable this scaling, including our approaches that rethink the modeling of large-scale structure surveys from the ground up. These developments are priming us for the next phase of scaling our methods to present-day and next-generation surveys while including parameterizations of galaxy-scale astrophysics and survey systematics. Already, the public release of our LtU-ILI inference framework is spurring community-wide adoption of the simulation-based inference techniques pioneered by our collaborators. This is evidenced by applications ranging from the recent weak lensing analyses of KiDS-1000 data to investigations of high-redshift galaxies with JWST. Looking ahead, I will outline our recent advances extending this modeling framework to enable joint analyses of large-scale structure and CMB data, specifically targeting datasets from ACT and the upcoming Simons Observatory. This work is paving the way for comprehensive, multi-probe cosmological inference at a scale previously thought impossible.