Abstract
Many different studies have shown that a wealth of cosmological information resides on small, nonlinear scales. Unfortunately, there are two challenges to overcome to utilize that information. First, we do not know the optimal estimator that will allow us to retrieve the maximum information. Second, baryonic effects impact that regime significantly and in a poorly understood manner. Ideally, we would like to use an estimator that extracts the maximum cosmological information while marginalizing over baryonic effects. In this work we show that neural networks can achieve that when considering some simple scenarios. We made use of data where the maximum amount of cosmological information is known: power spectra and 2D Gaussian density fields. We also contaminate the data with simplified baryonic effects and train neural networks to predict the value of the cosmological parameters. For this data, we show that neural networks can (1) extract the maximum available cosmological information, (2) marginalize over baryonic effects, and (3) extract cosmological information that is buried in the regime dominated by baryonic physics. We also show that neural networks learn the priors of the data they are trained on, affecting their extrapolation properties. We conclude that a promising strategy to maximize the scientific return of cosmological experiments is to train neural networks on state-of-the-art numerical simulations with different strengths and implementations of baryonic effects.
Introduction
Modern cosmology seeks to determine the fundamental parameters of our universe with unprecedented precision. While large-scale structures are well-modeled by linear theory, a significant amount of information is contained within small, nonlinear scales. However, extracting this information faces two primary hurdles: the lack of a known optimal summary statistic for small scales and the presence of complex, poorly understood baryonic effects (e.g., supernova feedback, gas cooling) that contaminate the signal. This paper investigates whether neural networks (NNs) can serve as optimal estimators that simultaneously maximize information extraction and marginalize over these baryonic uncertainties.
Methodology
To rigorously test the optimality of NNs, the authors employed "toy models" where theoretical performance limits, such as Fisher information or Maximum Likelihood Estimators (MLE), are analytically calculable.
Data Generation
The study utilized two primary data types:
- 1D Power Spectra: Modeled as power laws $P(k) = Ak^B$.
- 2D Gaussian Density Fields: Richer datasets allowing Convolutional Neural Networks (CNNs) to learn spatial patterns directly.
Three distinct scenarios were created to simulate varying levels of baryonic interference:
- AstroNone: A pure cosmological signal without baryonic effects.
- AstroDis: Baryonic effects introduce a discontinuity at a specific scale ($k_{pivot}$), effectively decoupling small-scale information from the underlying cosmology.
- AstroCon: Baryonic effects are present, but the power spectrum remains continuous at $k_{pivot}$, creating a physical link between "clean" large scales and "messy" small scales.
Key Results
Optimality and Automatic Marginalization
In the "clean" AstroNone scenario, NNs matched the performance of the MLE and the Fisher matrix limit. For 2D Gaussian fields, the NN achieved a Root Mean Square Error (RMSE) of 0.0113 for parameter $A$, nearly identical to the theoretical limit of 0.0111.
When baryonic effects were introduced, the NNs demonstrated an innate ability to identify and "mask" contaminated regions without explicit instruction. In the AstroDis case, while error increased due to information loss, the NN remained an unbiased estimator by successfully ignoring the "junk" data at small scales.
Extracting Buried Information
A significant finding occurred in the AstroCon scenario. The NN outperformed models that simply ignored small scales, successfully leveraging the continuity requirement to extract cosmological information "buried" within the baryon-dominated regime.
The Role of Priors
The study also highlighted that NNs learn the distribution (priors) of their training data. While this provides stability in low-signal regimes compared to MLE, it can lead to biases if the true physical parameters fall outside the training range. This underscores the importance of diverse and representative training sets.
Conclusion and Future Directions
This work provides a proof-of-principle that neural networks can function as optimal cosmological estimators capable of marginalizing over baryonic effects. While these results are based on Gaussian toy models, they suggest a powerful path forward. The next logical step is applying these techniques to complex, non-linear data from hydrodynamical simulations, such as the CAMELS suite, to develop a robust, "universal" cosmological estimator for real-world observations.
Summary updated