Linking galaxy properties to halo gas density profiles with interpretable machine learning • Daniele Sorini
Stellar and AGN feedback shape the distribution of gas from galaxies out to the intergalactic medium, but how this connects to galaxy properties remains uncertain. In this talk, I present a random forest algorithm that predicts the radial gas density profile within haloes from global properties of the central galaxy, including gas and stellar mass, star formation rate, and black hole mass and accretion rate. The model is trained on cosmological hydrodynamical simulations (EAGLE, IllustrisTNG, and Simba) and accurately recovers the simulated gas density profiles across a wide range of halo masses and redshifts. I will show how the predictions can be interpreted using statistical tools, including Sobol sensitivity analysis, to identify which galaxy properties are most closely linked to the gas distribution and what this reveals about feedback physics.
Spotting the Giants: Finding Galaxy Clusters in Noisy Weak-Lensing Convergence Maps • Gavin Leroy
Galaxy clusters are among the most massive bound structures in the Universe and are key probes of cosmology and large-scale structure formation. In weak gravitational lensing surveys, they can appear as localized features in convergence maps. However, their detection is complicated by projection effects, shape noise, filtering choices, and the intrinsically non-Gaussian structure of the cosmic matter field.
In this work, we investigate supervised machine-learning approaches for detecting galaxy clusters in filtered convergence maps, comparing convolutional U-Net and Vision Transformer models.
Trustworthy Digital Twins of AREPO AGN-Jet Simulations • Craig Bower
AGN jets are among the most powerful regulators in the Universe, yet each high-resolution AREPO simulation costs millions of core-hours, and every jet power demands a fresh run. We explore how a fast neural emulator can act as a digital twin of these simulations.
Our physics-informed neural field maps position, time, and Eddington ratio to hydrodynamic fields while enforcing physical constraints and enabling fast querying.
We also demonstrate how adaptive conformal inference provides reliable uncertainty estimates across different regimes.
AI and DiRAC Training Hub • Gokman
In DiRAC Training Academy, there are several JupyterHub charts that allow users run and test their codes. Porting all codes on those complex machines could be harder than practice. In this way, we design to help users how they can run and test a piece of their codes on those architechtures to see performance of functions/models easily. So, they do not have to run heavy profiler, tools and deal with all different software stack set up. By using containers and debuggers, we aim to help users/researchers how to use machine learning methods, hyper-parameter tuning in domain of scientific computing.
*This talk has been cancelled due to planing issue.*