Abstract
We demonstrate the potential of Deep Learning methods for measurements of cosmological parameters from density fields, focusing on the extraction of non-Gaussian information. We consider weak lensing mass maps as our dataset. We aim for our method to be able to distinguish between five models, which were chosen to lie along the $\sigma_8$ - $\Omega_m$ degeneracy, and have nearly the same two-point statistics. We design and implement a Deep Convolutional Neural Network (DCNN) which learns the relation between five cosmological models and the mass maps they generate. We develop a new training strategy which ensures the good performance of the network for high levels of noise. We compare the performance of this approach to commonly used non-Gaussian statistics, namely the skewness and kurtosis of the convergence maps. We find that our implementation of DCNN outperforms the skewness and kurtosis statistics, especially for high noise levels. The network maintains the mean discrimination efficiency greater than $85\%$ even for noise levels corresponding to ground based lensing observations, while the other statistics perform worse in this setting, achieving efficiency less than $70\%$. This demonstrates the ability of CNN-based methods to efficiently break the $\sigma_8$ - $\Omega_m$ degeneracy with weak lensing mass maps alone.
Introduction: The Challenge of Degeneracy
In modern cosmology, determining the fundamental parameters of our universe is a primary goal. Two critical parameters are $\Omega_m$, representing the total matter density, and $\sigma_8$, which describes the amplitude of matter fluctuations on scales of 8 $h^{-1}$ Mpc. However, standard observational techniques often face the "degeneracy" problem.
Traditional two-point statistics, such as the power spectrum, measure the correlation between points at various distances. While powerful, these statistics often cannot distinguish between a universe with high matter density and low fluctuations versus one with low density and high fluctuations. On a plot of $\sigma_8$ versus $\Omega_m$, these models fall along a curve where they appear statistically identical to second-order tools.
To break this degeneracy, researchers must look for "non-Gaussian" information—the structural signatures of the cosmic web, such as filaments, voids, and massive clusters. While higher-order statistics like skewness and kurtosis attempt to capture this, their performance degrades significantly in the presence of observational noise.
Methodology: Deep Learning for Mass Maps
This study proposes a Deep Convolutional Neural Network (DCNN) to classify weak lensing convergence maps—2D projections of the matter distribution inferred from the gravitational bending of light.
Network Architecture
The researchers implemented a custom DCNN featuring five convolutional layers followed by a fully connected layer with 1024 neurons. Unlike off-the-shelf architectures like VGG-16, this model was streamlined to prevent overfitting on the specific features of cosmological simulations while remaining deep enough to extract complex hierarchical patterns.
Noise-Adaptive Training (Curriculum Learning)
A significant hurdle in applying DCNNs to astronomical data is the high level of noise inherent in observations. Training a network directly on noisy data often fails because the complex optimization landscape prevents the model from converging on meaningful physical features.
To overcome this, the authors developed a "noise-ramping" strategy, a form of curriculum learning. The network is initially trained on noiseless maps. Once it reaches a predefined accuracy threshold ($T_a$), the noise level is incrementally increased. This allows the network to first learn the underlying physical structures before learning to identify them through the "fog" of observational noise.
Data and Simulations
The study utilized 12,500 independent convergence maps generated using the L-PICOLA fast N-body simulation code. Five distinct cosmological models were selected, all positioned along the $\sigma_8 - \Omega_m$ degeneracy line to ensure they had nearly identical power spectra. Gaussian noise was added to simulate two realistic observational scenarios:
- Space-based surveys: Low noise ($n_g = 100$ galaxies/arcmin²).
- Ground-based surveys: High noise ($n_g = 10$ galaxies/arcmin²).
Results and Discussion
The performance was evaluated using Mean Discrimination Efficiency (MDE)—the probability of correctly identifying the true model among the five candidates.
In the noiseless scenario, both the DCNN and traditional higher-order statistics (skewness and kurtosis) achieved near-perfect accuracy. However, as noise increased to levels typical of ground-based observations, a stark difference emerged. The DCNN maintained a robust MDE of over 85%, whereas the performance of skewness and kurtosis dropped to below 70%.
This suggests that the DCNN automatically learns "filters" that are more resilient to pixel-level noise than simple mathematical moments. Visualization of the network's internal feature maps indicated that it focuses on high-density peaks (massive clusters) and utilizes contrast-enhancing features to distinguish between models.
Conclusion and Future Directions
This work demonstrates that DCNNs can effectively break the $\sigma_8 - \Omega_m$ degeneracy by extracting non-Gaussian information that traditional statistics miss, especially in noisy data. While this study treated the problem as a discrete classification task, future work will involve transitioning to regression-based models for continuous parameter estimation. Furthermore, addressing the "simulation-to-real" gap—ensuring networks trained on simulations remain accurate when applied to real telescope data with complex systematics—remains a critical next step for the next generation of surveys like Euclid and LSST.
Summary updated