2025 workshop program
Machine learning-based emulator for large-volume semi-analytical galaxy formation models
The work
Abstract
We present progress toward the development of a machine learning-based emulator for the L-Galaxies semi-analytical galaxy formation model, applied to large-volume cosmological simulations. Our goal is to emulate the outputs of L-Galaxies within the MTNG (MillenniumTNG) dark matter only simulation suite, using the 740 Mpc box with updated cosmological parameters. As a proof of concept, we have trained a fully connected neural network on ∼300 realizations of the L-Galaxies model run on the original Millennium-I simulation (480 Mpc, Planck-I cosmology). The proof of concept reproduces key global galaxy properties, including the quenched galaxy stellar mass function. Initial results demonstrate a modest but promising performance. This work is part of a broader planned effort to build fast emulators and differentiable models for use in inverse modeling, Bayesian inference, and cosmological likelihood pipelines. Improved fidelity will require both a more robust neural architecture—potentially incorporating transformer-based models—and a significantly larger, higher-resolution training dataset. This poster will present the current state of the emulator, its capabilities, and our roadmap toward deployment on MTNG-scale simulations.