Deep Generative Modeling for Cosmological Fields

This repository presents a modular, research-level pipeline for simulating, compressing, and performing rigorous likelihood-based inference on high-dimensional Gaussian Random Fields (GRFs). The approach leverages state-of-the-art deep learning techniques—autoencoders and normalizing flows (RealNVP)—to enable scientific parameter estimation in cosmology and beyond.

---

Project Overview

Modern cosmological analysis often involves high-dimensional data (such as cosmic fields or sky maps) where direct likelihood evaluation is intractable. This project demonstrates how to:

• Generate synthetic cosmological fields with controlled parameters,
• Compress them using deep convolutional autoencoders into informative latent representations,
• Learn flexible probabilistic models (normalizing flows) for these latent spaces,
• Perform explicit, likelihood-based inference on the original physical parameters.

The resulting pipeline not only enables parameter recovery from simulated data but also provides a framework for deploying these techniques on real astronomical observations or other scientific fields.

---

Contents

• new_autoencoder.ipynb
  • End-to-end notebook for synthetic GRF data generation, autoencoder model training, validation, and visualization of reconstructions.
  • Modular code blocks for adjusting model architecture and dataset parameters.
• NVP_Flow:_conditional.ipynb
  • Implements conditional RealNVP flows in the autoencoder latent space.
  • Enables likelihood evaluation and parameter inference as a function of cosmological parameters.
  • Includes visualization routines for likelihood surfaces and parameter posteriors.
• NVP_flow.ipynb
  • Standalone demo of RealNVP on toy problems (e.g., two moons) to build intuition before applying to scientific data.
• grf_autoencoder.pth
  • Pretrained model weights for the autoencoder, ready for immediate use or fine-tuning.
• figs/
  • Collection of sample figures: reconstructions, loss curves, flow samples, and likelihood maps.

---

Key Strengths

• Synthetic Data with Physical Motivation:
  The code generates GRFs from parametric power spectra, mimicking real cosmological field statistics. This allows controlled benchmarking of inference pipelines.

• Modular Deep Learning Components:
  Easily swap or extend model architectures. Autoencoders are built for flexibility (depth, bottleneck size, activation), supporting rapid experimentation.

• Explicit Likelihood Evaluation:
  Conditional normalizing flows provide tractable likelihoods in compressed spaces—enabling rigorous, Bayesian-like parameter estimation.

• Visualization & Diagnostics:
  Notebooks include clear visual outputs: sample fields, reconstructions, latent space structure, flow-generated samples, and likelihood surfaces over parameter space.

• Fully Reproducible & Extensible:
  Synthetic data generation is integrated—no external datasets required. All hyperparameters and random seeds can be controlled for reproducibility.

---

How to Use

1. Clone the repository and install dependencies:

   pip install numpy scipy matplotlib torch scikit-learn tqdm
   

2. Run new_autoencoder.ipynb to:
   • Generate synthetic GRF data,
   • Train the convolutional autoencoder,
   • Visualize reconstructions and training curves.
3. Run NVP_Flow:_conditional.ipynb to:
   • Train a conditional normalizing flow in the latent space,
   • Evaluate and visualize likelihood surfaces over cosmological parameter space.
4. (Optional) Run NVP_flow.ipynb to experiment with RealNVP on toy datasets for intuition-building.
5. Browse the figs/ directory to see generated figures and outcomes.

The pipeline is designed to be run out-of-the-box; if data files are missing, they will be generated automatically.

---

---

Repository Structure

├── figs/
│   └── [Generated figures: reconstruction plots, likelihood maps, etc.]
├── grf_autoencoder.pth
├── new_autoencoder.ipynb
├── NVP_Flow:_conditional.ipynb
├── NVP_flow.ipynb
├── requirements.txt

---

Future Improvements

• More expressive generative models:
  Add deeper or alternative architectures (e.g., ResNet-based autoencoders, GLOW/MAF/Neural Spline Flows).

• Uncertainty quantification:
  Overlay credible contours on likelihood maps; integrate full Bayesian posteriors.

• Application to real datasets:
  Adapt data pipeline for telescope or simulation outputs; add data augmentation or physical systematics.

• Experiment tracking & reproducibility:
  Integrate MLflow or Weights & Biases for experiment management and hyperparameter sweeps.

• Comprehensive testing:
  Add unit tests for all model components and utility functions.

---

Contact

Mohammad Farhan Hassan
hassan.farhan7777@gmail.com

---

This project demonstrates the intersection of deep generative modeling and scientific inference—showcasing modern ML techniques applied to synthetic cosmological data, with extensibility to many domains in science and engineering.

---

Feel free to fork this repository or open issues for questions, suggestions, or collaborations!