The FLAMINGO project: From High-performance computing to simulating the whole Universe
Abstract:
The interpretation of data coming from cosmology surveys (such as the recently launched Euclid satellite) rely on comparison with accurate theoretical models including all the known relevant physical phenomena. The precision reached by modern instruments make this a extremely challenging task for numerical physicists who have to make use of some of the largest HPC facilities to run their calculations. In this talk, I will talk about the recently completed FLAMINGO project, a virtual twin to our own universe. This suite of simulations contains, among other runs, the largest cosmological calculation ever performed. I will introduce the key physics and cosmology question as well as cover some of the technical computational challenges and the solutions we implemented in the SWIFT cosmology code to overcome them.
Bio:
Mattieu Schaller is an assistant professor in numerical cosmology at the Lorentz Institute for theoretical physics and Leiden Observatory, where he works on the development and analysis of cosmological simulations. His research focuses the development of numerical simulation tools for cosmology and astrophysics, mainly on the SWIFT code and associated packages, as well as the preparation, running, and analysis of galaxy formation and cosmology simulations, such as the state-of-the-art EAGLE, SIBELIUS, FLAMINGO, and COLIBRE projects, which have been used in 100s of subsequent research studies around the world. The research in his group encompasses the low-level technical challenges of high-performance computing, the development of accurate numerical methods, and the construction of tools to interpret the simulated results and confront them to the observed Universe.
Summary
Pillars of modern physics
General Relativity: gravity
Quantum Field Theory
Standard Model of Particle Physics
Lambda-Cold Dark Matter (λCDM) model of universe
Cosmic microwave background: temperature variations in early universe
Can describe power spectrum using a 6-parameter model based on cosmological parameters
Cosmology tension: model’s predictions differ from astronomical measurements
Are we missing laws of physics or is it a measurement error?
Approach: simulation of the universe to predict how it should look like after evolving from different starting conditions
EAGLE: Evolution and Assembly of GaLaxies and their Environments: https://icc.dur.ac.uk/Eagle
Gravitational collapse of more dense regions evolved into
filaments and nodes of denser regions and
then into individual galaxies
Depending on the different initial conditions the distribution of this filament graph changes
Simulations that incorporate different proposals for dark matter make very different predictions
Few, massive particles vs Many, light particles
Details that we may want to model:
Initial conditions: known from CMB
Physical dynamics: gravity, hydrodynamics, magnetic fields, radiative transfer, cosmic rays
Constituents: dark energy, dark matter, normal matter, other details of matter (stars, planets, etc.)
Depending on level of detail this can go from easy to very hard on available supercomputers
Challenges of developing this simulation
Gravity is hard:
Very long-range: so need to exchange information among many different simulated objects
Always attractive: errors accumulate during the simulation (additive errors not compensated by subtraction)
All scales matter almost equally (hard to approximate)
Cheap calculation: inefficient use of available compute resources
Gas and stars affect the evolution of the universe
Galaxy is emitting hot gas
Powered by galaxy center region 1e3 times smaller
Powered by a central black hole that is 1e6 times smaller than that
Full resolution computation on an exascale computer would take 1e24 seconds = 1e6 times age of universe
Need software solutions to make this tractable:
Multiscale
Multi-grid
Particle splitting
(semi-)Lagrangian
Focus most compute on one single region
Load varies in both time and scale, to need to dynamically adapt numerical scheme based on current state of the simulation
Dynamic re-meshing
Task-based parallelism
But this means that the computer spends more time on management logic at the expense of less computation
Example: Astro-SPH
Task-based parallelism
Threads work on regions of space that are available
Modeling astrophysical details
Large scale universal structures can be fully simulated
Small scale structures need to be approximated via sub-grid models
cooling/heating of gas
Star formation
Enriching of gas from stars
Black holes
….
These models are approximate, so hard to capture their error or how it pollutes the dynamics of the larger simulation
Simple empirical parameter-fit models
Machine learning approach:
Run simulation of target phenomenon and train an ML model on its dynamics
ML model is approximate and very cheap to run
Use ML model to simulate small-scale phenomena
Observational approach: take observations of relevant phenomenon
FLAMINGO project: https://flamingo.strw.leidenuniv.nl/
Models many phenomena simultaneously: gravity, gases, stars, black holes
Calibration of sub-grid models
Run many simulations with different parameters for the sub-grid models
Use Gaussian processes to compute the probability distribution on the most likely values of these parameters
Validated against observational data
Fine-grained simulations not used for validation because the basic physics of stars, black holes, etc. are still being developed
Computation Cost:
42 days on 30k CPUs
31m CPI hours
145 MWh of power use
£20k cost
3.9T of CO2
Validation against observational data must account for the properties of telescopes, atmosphere, interference from the moon (affects the distribution of sky patches that could be observed), etc.
Outcome: predictions still differ from the observations, so more work to be done on understanding the physics that drive the universe
Debugging
For solvable/standard parts of the PDEs can compare to reference solutions or other simulations
To figure out which sub-grid models are most responsible for the error, need to look for which parts of the prediction are wrong
There are questions about whether the major laws (e.g. gravity) need revision. Simulation allows modeling of alternatives.