Presentasjon av masteroppgave: Eric Ludvigsen
Advertisement
«Fast Baryon Field Generation from Dark Matter Seeds via a 3D Deep Learning Model»
Abstract:
Producing large-scale cosmological simulations with both baryon and dark matter distributions is a computationally expensive process. The CAMELS project has produced existing (magneto)hydrodynamic simulations suitable for training machine learning algorithms and which can be used for data generation or parameter estimation. This data provides both raw training material, as well as a target for models to reproduce.
Producing a new machine learning model that is conditioned on a pure dark matter simulation and the parameters Ωm, σ8, ASN1, AAGN1, ASN2, AAGN2, and which can learn the statistical correlation between a dark matter ‘seed’ and a full baryon distribution. This model can then be used to predict complete baryon distributions much faster than running an exhaustive N-body magnetohydrodynamic simulation.
We start with an existing machine learning model, EMBER-2, which can perform field predictions of 2D baryon distributions from pure dark matter fields. We adapt the model to 3 dimensions, produce a pipeline that can accept CAMELS data and run large-scale training, and generate predictions with multiple trained models.
The model can visually and statistically reproduce large scale structure when compared to target fields, but is somewhat noisy when it comes to small scale structures, and performs better for matter than for energy and temperature fields. Results were verified by checking the power spectrum and magnitude distributions. For baryons and dark matter, we can produce power spectra that match to within 10% inside the length scales represented by a range in wave number k of about 0.3– 10 h Mpc−1. Very approximately, this should correspond to a maximum scale around 25 h−1 Mpc, and a minimum scale around 600 h−1 kpc.
This new machine learning model can be used to cheaply create new baryon distributions from dark matter seeds, which enables future work in parameter estimation and potentially testing theories of modified gravity with much faster iteration times than previously. We also propose some possible adjustments to further improve results and model performance.
Image caption: Overview figure showcasing simulated and emulated baryon fields at z = 1.8. Each pair of panels shows the simulated and emulated fields from
the FB30 simulation and the EMBER-2 model, respectively using the same colour scale. M. Bernardini et al., https://doi.org/10.1093/mnras/staf341
----------------------------------
More information: https://www.mn.uio.no/astro/studier/aktuelt/arrangementer/masterpresentasjoner/2026/Eric%20Ludvigsen.html
Abstract:
Producing large-scale cosmological simulations with both baryon and dark matter distributions is a computationally expensive process. The CAMELS project has produced existing (magneto)hydrodynamic simulations suitable for training machine learning algorithms and which can be used for data generation or parameter estimation. This data provides both raw training material, as well as a target for models to reproduce.
Producing a new machine learning model that is conditioned on a pure dark matter simulation and the parameters Ωm, σ8, ASN1, AAGN1, ASN2, AAGN2, and which can learn the statistical correlation between a dark matter ‘seed’ and a full baryon distribution. This model can then be used to predict complete baryon distributions much faster than running an exhaustive N-body magnetohydrodynamic simulation.
We start with an existing machine learning model, EMBER-2, which can perform field predictions of 2D baryon distributions from pure dark matter fields. We adapt the model to 3 dimensions, produce a pipeline that can accept CAMELS data and run large-scale training, and generate predictions with multiple trained models.
The model can visually and statistically reproduce large scale structure when compared to target fields, but is somewhat noisy when it comes to small scale structures, and performs better for matter than for energy and temperature fields. Results were verified by checking the power spectrum and magnitude distributions. For baryons and dark matter, we can produce power spectra that match to within 10% inside the length scales represented by a range in wave number k of about 0.3– 10 h Mpc−1. Very approximately, this should correspond to a maximum scale around 25 h−1 Mpc, and a minimum scale around 600 h−1 kpc.
This new machine learning model can be used to cheaply create new baryon distributions from dark matter seeds, which enables future work in parameter estimation and potentially testing theories of modified gravity with much faster iteration times than previously. We also propose some possible adjustments to further improve results and model performance.
Image caption: Overview figure showcasing simulated and emulated baryon fields at z = 1.8. Each pair of panels shows the simulated and emulated fields from
the FB30 simulation and the EMBER-2 model, respectively using the same colour scale. M. Bernardini et al., https://doi.org/10.1093/mnras/staf341
----------------------------------
More information: https://www.mn.uio.no/astro/studier/aktuelt/arrangementer/masterpresentasjoner/2026/Eric%20Ludvigsen.html
Advertisement
Where is it happening?
Institute of Theoretical Astrophysics, University of Oslo, Sognsveien 77B, 0855 Oslo, Norge, Oslo, Norway
Event Location & Nearby Stays:
Know what’s Happening Next — before everyone else does.
Host or PublisherInstitute of Theoretical Astrophysics, University of Oslo
