2026 workshop program
Learning to Jet: A 3-D UNet Enabled Subgrid Model for AGN Jet Feedback in Cosmological Simulations
The work
Abstract
Active galactic nuclei (AGN) are crucial components in quenching star formation within galaxies, shaping their long-term evolution. While current cosmological simulations reproduce some observed galaxy properties, they fail to match others, especially at high redshift, suggesting that existing AGN feedback prescriptions are still missing essential physics. Here, we introduce a novel method for modeling AGN jet feedback: a diffusion model with a 3D U-Net architecture, trained to predict AGN jet evolution across time intervals comparable to those used in cosmological simulations. Our model is trained on pairs of coarse-grained, high-resolution AGN jet simulations that resolve the turbulent medium within the central 1 kpc of the host galaxy. We aim to reproduce the spatial distribution of eight key physical parameters—number density, internal energy, and six velocity components—to evolve and reconstruct the full energy distribution. We find that both model size, dimension, and the choice of time-step separation, $\Delta t$, are critical hyperparameters that influence model fidelity. Our final model is able to predict jet evolution over large time-steps, with little compounding error over repeated applications of the model to its own output.