Impact of numerical solver differences across simulation codes on field-level results

Determine the extent to which differences in numerical implementations across cosmological simulation codes—such as TreePM gravity solvers (e.g., Arepo, Gadget), adaptive-mesh-refinement with FFT gravity (Enzo), the Fast Multipole Method (PKDGRAV3), particle–mesh approaches (CUBEP3M), and multipole approximations (Abacus)—affect the field-level results used for machine-learning-based inference of dark-matter and astrophysical parameters.

Background

The DREAMS project leverages large suites of hydrodynamical and N-body simulations, along with machine-learning methods, to infer dark matter properties while marginalizing over astrophysical and cosmological uncertainties. However, different simulation codes employ distinct numerical techniques for gravity and hydrodynamics (e.g., TreePM, AMR with FFT, Fast Multipole, particle–mesh, and multipole approximations), which can introduce subtle numerical deviations in the simulated matter and galaxy fields.

Prior studies have shown that machine-learning models trained on one code may not generalize well to outputs from another, suggesting that code-specific numerical details can impact downstream inference. While training across multiple codes may improve robustness, the magnitude and nature of these numerical effects on field-level results remain to be quantified rigorously.