TL;DR
This paper introduces a novel approach using Graph Neural Networks (GNNs) to infer cosmological parameters ($Ω_m$ and $σ_8$) directly from dark matter halo catalogs. Remarkably, these GNNs demonstrate strong robustness, accurately predicting cosmological parameters even when trained on simulations from one code and tested on data from entirely different simulation codes and even complex hydrodynamic simulations.
What Problem Does This Paper Address?
Cosmologists aim to precisely measure fundamental parameters like the matter density ($Ω_m$) and the amplitude of matter fluctuations ($σ_8$) to understand the Universe's composition and evolution. Traditionally, this involves compressing complex observational data into "summary statistics" and then comparing these to theoretical predictions. However, for non-linear structures like dark matter halos, the optimal summary statistics are unknown, leading to a loss of information. Furthermore, existing machine learning approaches often struggle with "robustness" – failing when applied to data generated by different simulation codes due to subtle differences in their numerical methods and physics implementations. This paper tackles the challenge of developing a robust, field-level inference method that can extract cosmological information from dark matter halo distributions without relying on pre-defined summary statistics and can generalize across diverse simulation outputs.
Methods
The core of this work lies in the application of Graph Neural Networks (GNNs). GNNs are well-suited for this task because they can naturally handle the irregular structure of halo catalogs, treating each halo as a node and their spatial relationships as edges. The authors designed their GNNs to be permutationally, translationally, and rotationally invariant, ensuring that the model's predictions are independent of the order of halos, their absolute positions, or their orientation.
Key methodological aspects include:
- Halo Graphs: Halo catalogs are converted into graph structures where halos are nodes. Edges connect halos within a certain linking radius.
- Node and Edge Features: Node features can include halo properties like mass, velocity, concentration, and maximum circular velocity. Edge features encode the relative positions and orientations of connected halos, crucial for enforcing symmetries.
- Message Passing Layers: The GNN architecture uses multiple "message passing" layers where nodes exchange information with their neighbors, iteratively updating their features.
- Likelihood-Free Inference: The GNNs are trained to predict the mean and standard deviation of the cosmological parameters, bypassing the need for an explicit likelihood function.
- Robustness Testing: A critical aspect is testing the trained models on halo catalogs generated by a wide array of N-body and hydrodynamic simulation codes, often with different cosmologies and initial conditions.
Datasets
The study utilizes a large suite of cosmological simulations:
- Training Data: Halo catalogs generated from the Gadget N-body simulation code were used for training.
- Testing Data: The models were rigorously tested on halo catalogs from thousands of simulations generated by:
- N-body codes: Abacus, CUBEP³M, Enzo, PKDGrav3, and Ramses.
- Hydrodynamic codes (CAMELS project): Astrid, IllustrisTNG, Magneticum, and SIMBA.
These simulations cover a range of cosmologies and, for hydrodynamic simulations, different astrophysical subgrid physics implementations. The simulations are typically run in small, periodic volumes of $(25~h^{-1}{\rm Mpc})^3$.
Key Insights and Ideas
- GNNs as Universal Cosmological Probes: The success in inferring cosmological parameters from diverse simulation codes suggests that GNNs can learn fundamental relationships in the cosmic web that are largely independent of specific simulation implementations.
- Robustness is Achievable: The paper demonstrates that robust field-level inference is possible, a significant step beyond previous methods that often failed to generalize.
- Clustering is Key for $Ω_m$: The study highlights that the spatial distribution (clustering) of halos provides the dominant information for inferring $Ω_m$, more so than individual halo velocities.
- $σ_8$ Inference is More Sensitive: Inferring $σ_8$ proves more challenging and requires more stringent conditions (e.g., larger halos) for robustness, especially when dealing with hydrodynamic simulations. This suggests $σ_8$ is more sensitive to baryonic physics and simulation details.
- Trade-off Between Precision and Robustness: Including more detailed halo properties (like concentration and maximum circular velocity) can increase the precision of the inference but often at the cost of robustness, as these properties are more sensitive to simulation-specific details.
Key Results
- Accurate $Ω_m$ Inference: The GNNs can infer $Ω_m$ with a mean relative error of $\sim6\%$ when using halo positions and velocities, and this accuracy is maintained across five different N-body codes and four hydrodynamic codes.
- Robust $Ω_m$ Inference: The model trained on $Ω_m$ using Gadget simulations is surprisingly robust and performs well even on hydrodynamic simulations, indicating it captures universal physics.
- Accurate $σ_8$ Inference: Using halo positions and masses, the GNN infers $σ_8$ with a mean relative error of $\sim5.9\%$.
- Robustness for $σ_8$ is Conditional: The robustness of the $σ_8$ inference is more sensitive to halo mass, requiring halos with at least $\sim300$ particles for N-body simulations. It is not robust for hydrodynamic simulations.
- Information from Clustering: The GNNs extract significantly more information for $Ω_m$ inference compared to traditional methods like the halo power spectrum.
Limitations and Future Work
- Robustness to Hydrodynamics: While the $Ω_m$ model shows remarkable robustness to hydrodynamic simulations, the $σ_8$ model does not, suggesting that baryonic physics and subgrid models have a stronger impact on $σ_8$.
- Sensitivity to Halo Properties: Including detailed halo properties like concentration and maximum circular velocity improves precision but breaks robustness, as these are more sensitive to simulation specifics.
- Redshift and Halo Finder Dependence: The models trained at $z=0$ do not extrapolate well to higher redshifts or to halo catalogs generated by different halo finders (e.g., SUBFIND vs. ROCKSTAR).
- Future Work: The authors suggest investigating the impact of including only central galaxies (instead of all halos) for more robust inference, exploring redshift-space distortions, and using larger simulation volumes to improve precision and potentially robustness. They also plan to use interpretability tools to understand how the GNNs achieve their robust predictions.
Summary updated