TL;DR
This paper introduces a novel approach using Graph Neural Networks (GNNs) to infer cosmological parameters directly from dark matter halo catalogs. Remarkably, these GNNs demonstrate robust performance across simulations from different codes and even complex hydrodynamic simulations, achieving ~6% accuracy for key parameters like Ωm and σ8.
What Problem Does This Paper Address?
Cosmologists aim to precisely measure fundamental parameters of the Universe, such as the matter density (Ωm) and the amplitude of matter fluctuations (σ8), by analyzing the distribution of matter. Traditionally, this involves compressing complex observational data into "summary statistics" and then comparing these to theoretical predictions. However, for the non-linear scales where most information resides, the optimal summary statistics are unknown, leading to a loss of information. Furthermore, machine learning models trained on simulations often struggle to generalize to data from different simulation codes due to subtle differences in numerical methods and physics implementations. This paper tackles the challenge of developing a robust inference method that can extract cosmological information from halo catalogs across diverse simulation setups, overcoming the limitations of traditional summary statistics and improving model generalizability.
Methods
The core of this work lies in the application of Graph Neural Networks (GNNs). GNNs are particularly well-suited for this task because they can naturally handle the irregular structure of halo catalogs and are designed to be invariant to permutations, translations, and rotations. The authors construct "halo graphs" where halos are nodes and edges connect halos within a certain linking radius. These graphs are then fed into GNNs trained to predict cosmological parameters. The key innovation is the focus on "field-level inference," meaning the GNN learns directly from the spatial arrangement and properties of halos, rather than relying on pre-defined summary statistics. The models are designed to be robust by training on a diverse set of simulations and by generating halo catalogs with varying number densities.
Datasets
The study utilizes a vast collection of halo catalogs generated from thousands of N-body and hydrodynamic simulations.
- N-body Simulations: Catalogs were generated from simulations run with six different N-body codes: Abacus, CUBEP³M, Enzo, Gadget, PKDGrav3, and Ramses. These simulations cover a range of cosmologies and initial conditions.
- Hydrodynamic Simulations: Catalogs were also generated from four state-of-the-art hydrodynamic simulation suites: Astrid, IllustrisTNG, Magneticum, and SIMBA. These simulations incorporate complex baryonic physics.
The primary training set consists of halo catalogs from Gadget simulations, while the models are tested on catalogs from all other simulation codes.
Key Insights and Ideas
- GNNs as Optimal Estimators: The study demonstrates that GNNs can act as powerful, data-driven estimators for cosmological parameters, surpassing traditional methods like the power spectrum in extracting information from halo distributions.
- Robustness is Achievable: A major breakthrough is the demonstration of robust inference across different N-body codes, even when they employ distinct numerical techniques. This suggests that the underlying physical relations learned by the GNNs are fundamental and not overly sensitive to simulation specifics.
- Hydrodynamics Don't Break Ωm Inference: Surprisingly, the model trained to infer Ωm remains robust even when tested on hydrodynamic simulations, which include complex baryonic physics. This indicates that the GNN can disentangle cosmological signals from baryonic effects for this parameter.
- σ8 is More Sensitive: In contrast, inferring σ8 proves more challenging and less robust, particularly when dealing with hydrodynamic simulations or when using halo properties like concentration and maximum circular velocity, which are more sensitive to simulation details.
- Trade-off Between Accuracy and Robustness: Including more detailed halo properties (like concentration and Vmax) can improve prediction accuracy on a single simulation but significantly degrades robustness across different simulation codes. This highlights the importance of choosing features that are well-converged across various simulation methods.
Key Results
- High Accuracy for Ωm and σ8: The GNNs achieve a mean relative error of approximately 6% for both Ωm and σ8 when trained on halo positions and velocities (for Ωm) or positions and masses (for σ8).
- Exceptional Robustness for Ωm: The Ωm inference model demonstrates remarkable robustness, accurately predicting Ωm across halo catalogs from five different N-body codes and four different hydrodynamic simulation suites.
- Partial Robustness for σ8: The σ8 inference model shows robustness across N-body codes but only for halos with a sufficient number of particles (~300+). It is not robust to hydrodynamic simulations.
- Information Source: For Ωm inference, the study finds that halo clustering provides the dominant source of information, rather than halo velocities alone.
Limitations and Future Work
- Robustness of σ8: The inference of σ8 is found to be less robust than that of Ωm, especially when dealing with hydrodynamic simulations or when using halo properties like concentration and maximum circular velocity. This suggests that σ8 is more sensitive to baryonic physics and simulation-specific details.
- Halo Finder Dependence: The model is not robust when tested on halo catalogs generated by a different halo finder (SUBFIND) compared to the one used for training (ROCKSTAR), indicating that halo finder differences can impact inference.
- Redshift Dependence: The models trained at redshift z=0 are unable to accurately predict cosmological parameters at higher redshifts (z=1), as expected, since redshift significantly alters halo clustering.
- Future Work: The authors suggest exploring the use of central galaxies instead of dark matter halos for more robust inference, investigating the impact of redshift-space distortions, and repeating the analysis with larger simulation volumes to better understand super-sample covariance effects. They also plan to use interpretability tools to understand how the GNNs are making their predictions.
Summary updated