Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BaryonBridge: Stochastic Interpolant Model for Fast Hydrodynamical Simulations

Benjamin Horowitz, Carolina Cuesta-Lazaro, Omar Yehia

TL;DR

BaryonBridge introduces a conditional stochastic interpolant framework to rapidly map dark matter density fields from fast particle-mesh simulations to baryonic fields relevant for Ly-$\alpha$ forest observables. The method conditions on cosmological and astrophysical parameters and uses a 3D U-Net to learn the drift in an SDE that connects $\delta_{\text{DM}}$ to $\delta_{\text{baryons}}$, enabling end-to-end differentiable field-level inference. Trained on CAMELS IllustrisTNG-LH and validated on the larger TNG50 volume, it achieves high fidelity in $N_{\mathrm{HI}}$, $T$, and Ly-$\alpha$ flux statistics, with errors typically below a few percent for relevant $k$ ranges and strong cross-correlations up to several $h\mathrm{Mpc}^{-1}$. The approach yields substantial speedups (e.g., ~7 GPU-min for inference; ~96 GPU-hours for training) and supports fast mock generation and dynamical forward modeling, offering a practical pathway for next-generation survey analyses and field-level inferences.

Abstract

Constructing a general-purpose framework for mapping between dark matter simulations and observable hydrodynamical simulation outputs is a long-standing problem in modern astrophysics. In this work, we present a new approach utilizing stochastic interpolants to map between cheap fast particle mesh simulations and baryonic quantities in three dimensions, requiring a total of 7 GPU minutes per 256^3 grid size simulation. Using the CAMELS multifield dataset, we are able to condition our mapping on both cosmological and astrophysical properties. We focus this work on hydrodynamical quantities suitable for Lya observables finding excellent agreement up to small spatial scales, k ~ 10.0 (h^(-1) Mpc) at z=2.0, for Lya flux statistics. Our approach is fully convolutional, allowing training on comparatively small volumes and application to larger volumes, which was tested on TNG50.

BaryonBridge: Stochastic Interpolant Model for Fast Hydrodynamical Simulations

TL;DR

BaryonBridge introduces a conditional stochastic interpolant framework to rapidly map dark matter density fields from fast particle-mesh simulations to baryonic fields relevant for Ly- forest observables. The method conditions on cosmological and astrophysical parameters and uses a 3D U-Net to learn the drift in an SDE that connects to , enabling end-to-end differentiable field-level inference. Trained on CAMELS IllustrisTNG-LH and validated on the larger TNG50 volume, it achieves high fidelity in , , and Ly- flux statistics, with errors typically below a few percent for relevant ranges and strong cross-correlations up to several . The approach yields substantial speedups (e.g., ~7 GPU-min for inference; ~96 GPU-hours for training) and supports fast mock generation and dynamical forward modeling, offering a practical pathway for next-generation survey analyses and field-level inferences.

Abstract

Constructing a general-purpose framework for mapping between dark matter simulations and observable hydrodynamical simulation outputs is a long-standing problem in modern astrophysics. In this work, we present a new approach utilizing stochastic interpolants to map between cheap fast particle mesh simulations and baryonic quantities in three dimensions, requiring a total of 7 GPU minutes per 256^3 grid size simulation. Using the CAMELS multifield dataset, we are able to condition our mapping on both cosmological and astrophysical properties. We focus this work on hydrodynamical quantities suitable for Lya observables finding excellent agreement up to small spatial scales, k ~ 10.0 (h^(-1) Mpc) at z=2.0, for Lya flux statistics. Our approach is fully convolutional, allowing training on comparatively small volumes and application to larger volumes, which was tested on TNG50.
Paper Structure (10 sections, 5 equations, 3 figures, 1 table)

This paper contains 10 sections, 5 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Simulations
  3. CAMELS and TNG50
  4. Particle Mesh Simulations
  5. Mapping to Ly-$\alpha$ Forest
  6. Stochastic Interpolant Models
  7. Results
  8. Inference over larger volumes: TNG50
  9. Parameter variations: CAMELS Boxes
  10. Discussion and Conclusion

Figures (3)

  • Figure 1: We show the input density FastPM generated density field and generated neutral hydrogen ($N_{HI}$), temperature (T), and inferred Ly-$\alpha$ flux from our BaryonBridge model compared to the simulated true result in the TNG50 simulation.
  • Figure 2: Comparison of redshift space Ly-$\alpha$ flux for both the original TNG50 simulation and that mapped via BaryonBridge code. We also compare to an analytical approximation given via the Flucuating Gunn-Peterson Approximation (FGPA) on the original dark matter field.
  • Figure 3: Power spectra statistics corresponding to the models in Tab. \ref{['param_table']}. The shaded region shows the variance of independent samples drawn from the same initial conditions; i.e. a model prediction of uncertainty in the underlying mapping.