Baryon Painting Optimization

Associated publication: Kéruzoré, F. et al. (2022), OJAp 6, 43.

This project established the foundation for the ML-based surrogate approach later extended in picasso, starting from a classical, parametric baryon model and rigorously benchmarking it as a surrogate for expensive hydrodynamic simulations.

The key experimental setup was the Borg Cube: a matched pair of cosmological simulations run from identical initial conditions - one gravity-only (fast, cheap), one full hydrodynamics (slow, expensive, physically complete). This paired dataset provides a ground-truth benchmark for evaluating any surrogate model: by applying the model to the gravity-only run and comparing its predictions to the hydrodynamic run for the same halos, we can measure the surrogate’s accuracy without any confounding factors.

I implemented the baryon model on HACC gravity-only outputs and optimized its parameters against the hydrodynamic ground truth using large-scale data processed on HPC systems at the Argonne Leadership Computing Facility (ALCF). The resulting surrogate achieves few-percent accuracy on gas density and pressure predictions across a redshift range \(z \in [0, 2]\), and reproduces the mass-tSZ scaling relation with only a small additional scatter relative to the full hydrodynamic simulation.

This validates the surrogate paradigm: gravity-only simulations, augmented with a learned or calibrated baryon model, can substitute for far more expensive hydrodynamic runs in contexts where large simulation volumes and high throughput matter more than per-object precision.