Technical Summary: Machine-learning Applications for Weak-lensing Cosmology

Problem Statement

Weak gravitational lensing serves as a premier probe for mapping the total matter distribution in the Universe, independent of the physical state of matter, thereby offering a direct window into dark matter and large-scale structure formation. However, extracting cosmological information from weak-lensing data faces significant challenges:

1. Non-Gaussianity: The matter distribution on small scales is highly nonlinear and hierarchical. Traditional analyses relying on two-point correlation functions (or power spectra) fail to capture the full non-Gaussian information inherent in the cosmic web.
2. Modeling Limitations: Conventional cosmological inference often assumes a Gaussian likelihood for summary statistics. This assumption is questionable for higher-order statistics, and constructing accurate analytic models for the mean and covariance of these statistics is computationally prohibitive or theoretically intractable.
3. Observational Noise: Weak lensing signals are contaminated by "shape noise"—the intrinsic ellipticity of galaxies—which is significantly larger than the shear signal. This noise obscures diffuse mass components, limiting the recovery of large-scale structures in mass maps.
4. Generative Bottlenecks: Robust cosmological inference and blind analysis challenges require vast numbers of realistic mock datasets. Traditional ray-tracing simulations based on $N$-body or hydrodynamical simulations are computationally expensive, limiting the number of realizations available for error estimation and likelihood-free inference.

Methodology

The chapter reviews the application of machine learning (ML) techniques to address these challenges, categorizing them into three primary domains: data-driven feature extraction, noise reduction, and generative modeling.

1. Data-Driven Feature Extraction and Inference

Instead of relying on handcrafted summary statistics (e.g., peak counts, Minkowski functionals), ML methods, particularly Convolutional Neural Networks (CNNs), are employed to learn optimal summaries directly from convergence maps ($\kappa$).

• Cosmological Parameter Inference: CNNs are trained to map observed convergence maps directly to cosmological parameters (e.g., $\Omega_{m0}, \sigma_8$). These networks can capture subtle features, such as gradients around convergence peaks, that traditional statistics miss.
• Likelihood-Free Inference (LFI): To bypass the need for explicit likelihood models (Eq. 1.15), neural density estimators are used. These models learn the posterior distribution $Prob(q|D_{obs})$ by forward-simulating mock data vectors at various parameter values. This approach handles non-Gaussian likelihoods and avoids the computational cost of sampling high-dimensional hierarchical models.
• Out-of-Distribution (OoD) Detection: Neural density estimators can be repurposed to detect when observed data falls outside the validity domain of the simulator, providing a diagnostic for simulator accuracy.

2. Denoising Lensing Mass Maps

Deep learning is applied to separate the gravitational signal from shape noise.

• Formulation: The problem is framed as estimating the conditional probability $Prob(\kappa_{signal} | \kappa_{obs})$, where $\kappa_{obs} = \kappa_{signal} + \kappa_{noise}$.
• Architectures:
  • Conditional Generative Adversarial Networks (cGANs): Trained on paired datasets to produce a single plausible realization of the noiseless map.
  • Diffusion Models (Score-based Generative Models): These models offer a principled approach to learning the conditional distribution, preserving Gaussian statistics on large scales and allowing for the generation of multiple realizations to quantify reconstruction uncertainty.
  • Transformers: Proposed to handle survey masks and curved-sky effects more naturally than standard CNNs designed for flat images.

3. Deep Generative Models

To overcome the computational cost of ray-tracing simulations, deep learning models are used to generate realistic weak-lensing mass maps.

• Generative Adversarial Networks (GANs): Can rapidly produce realistic convergence fields after training. Variants include conditional GANs (conditioned on cosmological parameters) and CycleGANs (for unpaired translation and larger sky coverage).
• Diffusion Models: Grounded in a well-defined probabilistic framework, these models offer improved training stability and sample diversity compared to GANs, though at a higher computational cost during generation.
• Normalizing Flows: Provide the distinct advantage of exact evaluation of the data-generating probability density, though they require invertible mappings and careful architectural tuning.

Key Contributions and Findings

The paper synthesizes a decade of research, highlighting several key findings:

• Superiority of ML Summaries: Studies demonstrate that CNN-based summaries can improve cosmological constraints by a factor of $\sim 5$ over traditional two-point statistics in noiseless simulations, and still outperform power-spectrum analyses in the presence of realistic shape noise when appropriate smoothing scales are used.
• Information Content: Saliency analyses reveal that CNNs extract information from high-convergence regions and subtle gradients often missed by standard statistics.
• Denoising Efficacy: Deep learning approaches (cGANs and diffusion models) have successfully recovered small galaxy clusters and large-scale structure patterns that are otherwise hidden by shape noise in surveys like the Hyper Suprime-Cam (HSC) and Dark Energy Survey (DES).
• Generative Efficiency: Neural network-based generative models can produce realistic convergence fields with significantly fewer trainable parameters (e.g., $O(10^3)$ vs. $O(10^7)$ in conventional CNNs) and can generate vast numbers of realizations for error estimation, facilitating blind analysis challenges.
• Validation of LFI: Likelihood-free inference has been shown to provide credible intervals comparable to traditional methods, with strategies like pre-training on "cheap" simulations followed by fine-tuning accelerating convergence.

Significance and Future Outlook

The paper positions machine learning not merely as an alternative tool but as a necessary evolution for the next generation of weak-lensing cosmology.

• Field-Level Inference: The ultimate goal is field-level inference—estimating parameters directly from raw data without intermediate compression. ML generative models are identified as the key enabler for extracting information from highly nonlinear structures in a lossless manner.
• Interpretability and Efficiency: The author advocates for the development of "lightweight" neural networks that incorporate physical symmetries (homogeneity and isotropy) to improve interpretability and reduce computational overhead.
• 3D Reconstructions: Future work aims to extend mass mapping from 2D projections to 3D reconstructions of large-scale structures, leveraging the flexibility of deep learning to handle the complexities of deprojection.

The chapter concludes that while challenges regarding the physical interpretability of "black-box" models and the robustness of training on limited datasets remain, the integration of ML is poised to significantly enhance the scientific return of current and upcoming weak-lensing datasets.