DeepCHART

Deep learning for Cosmological Heterogeneity and Astrophysical Reconstruction via Tomography

This project implements a 3D tomographic reconstruction pipeline using Ly(\alpha) forest skewers and galaxy positions as input to reconstruct the dark matter (DM) density field. It is built using PyTorch and leverages a variational autoencoder (VAE) with a U-Net-based 3D convolutional neural network.

---

📦 Project Structure

DeepCHART/
├── config.py              # Configuration for training, model, and paths
├── dataset.py             # Custom PyTorch Dataset class for tau, galaxy, DM
├── model.py               # 3D VAE-UNet model with anisotropic kernels
├── train.py               # Training pipeline
└── README.md              # Project description

---

🧠 Core Features

• Two-channel input: sparse Ly(\alpha) forest and galaxy CIC fields
• Output: log-scaled 3D dark matter density
• Anisotropic volume support: reconstructs volumes of shape (N_\text{section} \times N_\text{section} \times Z_\text{section})
• Realistic noise model: SNR-based Gaussian noise injection in the flux
• Modular VAE-UNet: residual blocks, skip connections, and latent sampling

---

📐 Volume Specifications

• Input & output volumes are extracted as:

  shape = (N_SECTION, N_SECTION, Z_SECTION)  # typically 96 x 96 x 288

• Sliced from parent cubes of size N = 128 (configurable)

---

🚀 Getting Started

1. Configure your setup

Edit config.py to set:

• Data paths
• Network hyperparameters
• Volume shape (N_SECTION, Z_SECTION)
• SNR distribution and augmentation parameters

2. Run training

python train.py

3. Outputs

• Model checkpoint: model_tomography_...pth
• Training log: Output_F_galaxy_DM_...txt
• Loss plot: training_history.png

---

🔧 Requirements

Minimal requirements :

numpy
scipy
torch
matplotlib

---

📄 Citation

If you use this code, please consider citing:

Maitra et al. 2025, DeepCHART: Mapping the 3D dark matter density field from Lyα forest surveys using deep learning, arXiv:2507.00135

---