This project explores whether machine learning can predict the optical depth (τ) of the early universe from kSZ (kinetic Sunyaev-Zel'dovich) heat maps, using a dataset of 1,000 simulations provided by Dr. Paul La Plante. These simulations model various reionization scenarios and generate corresponding τ values and 2D kSZ maps of the Cosmic Microwave Background (CMB).
The τ value summarizes how much ionized gas scattered CMB photons along their journey from the early universe. Since the relationship between spatial kSZ features and τ is complex and nonlinear, we trained a neural network to learn this mapping and estimate τ values directly from kSZ maps.
- Source: Simulations by Dr. Paul La Plante
- Size: 1,000 reionization scenarios
- Each simulation includes:
- kSZ Heat Map: A 1024×1024 image of temperature shifts in the CMB
- Tau (τ) Value: Optical depth to the CMB
- Model Type: Semi-numeric "inside-out" reionization model
- Data loading and normalization
- Custom PyTorch dataset object
- Neural network architecture (CNN-based)
- CUDA acceleration for GPU training
- Training/validation split
- Model evaluation and prediction
- Sharp drop in loss after epoch 1
- No divergence between traiing and validation loss
- No signs of overfitting
- Model performs best on mid-range τ values (~0.4–0.5)
- Underestimates high τ and overestimates low τ
- Clear signs of regression to the mean
- Overestimation at low τ, underestimation at high τ
- Slight skew toward underestimating τ
- Most errors centered around zero
Successfully built a neural network to predict τ from simulated kSZ maps. Model captures trends but shows systematic bias toward the mean. Future work could include more complex models or better feature extraction techniques.