One of the most exciting results of recent times is a seeming tension between the values of the cosmological parameters inferred from the distant and very young universe and from the nearby universe. This possible tension in the measured parameters could be telling us that the cosmological standard model is incomplete. To get more accurate values for the parameters to see if the tension is real we not only need bigger and better telescopes and surveys, but also better models of the universe to compare to the data.
We require complex and very large hydrodynamic simulations that predict the distribution of gas, stars, black holes, neutrinos, and dark matter starting from just after the Big Bang to the present. For the large-volume simulations of the FLAMINGO suite of simulations, the biggest cosmological hydrodynamical simulation ever run, we have a particle resolution of about a billion solar masses, which implies that the stars in the Milky Way would be represented by about ten particles.
One of the main open questions is whether important processes in galaxy formation, specifically supernovae and the energy released by supermassive black holes, have a big impact on the inference of the cosmological parameters. Because our simulations are not able to resolve these processes in detail, they must be put in by hand as sub-resolution (sub-grid) models. While these sub-grid models model processes that play out below the resolution of the simulation they do impact scales that the simulations can resolve.
In our paper we use a new, faster and more robust method to calibrate the sub-grid models supernovae and black holes. For this calibration the predictions of the simulation are compared with observational data. To robustly fit the observations a large number of evaluations of the simulations are needed. This is where we use machine learning to speed up this process. To calibrate the sub-grid physics, we run 32 smartly sampled simulations and use machine learning to create an emulator that gives us the simulation prediction for the observables of interest in a fraction of a second, a speedup of more than a million. These emulators can then be put directly into a fitting pipeline, to find the best possible calibration of the sub grid physics. Because of using this novel method, the FLAMINGO simulations provide the most robust and accurate prediction for the effect of supernovae and black holes on our cosmological inference.
For more detail you can read the full paper here or visit the FLAMINGO web page:
Kugel et al 2023,MNRAS,526,6103 https://academic.oup.com/mnras/article/526/4/6103/7291940 and https://flamingo.strw.leidenuniv.nl/
The results of the new calibration method. The emulator, in blue, is able to accurately predict the simulation, in red, and was accurately fit to the data, in black.
