AstrophysicsPyTorchPhysics-Guided NNResearch
PREML
Photometric Redshift Estimation using Machine Learning
Project Dossier
My Final Year Project applying deep learning to redshift estimation. Redshift estimation is key in astrophysics to measure cosmic distances and understand the universe’s expansion. Traditional spectroscopic methods are accurate but not scalable. Our project tackles this using machine learning.
Technical Architecture
We developed a Physics-Guided Neural Network (PGNN) that processes both photometric magnitudes and galaxy images. The architecture consists of:
- MLP Head: Processes tabular photometric inputs.
- CNN Head: Extracts features from galaxy imagery.
- Cross-Attention: Fuses features from both heads for robust representation.
- Bayesian NN: Provides uncertainty estimation alongside predictions.
- Physics Loss: Incorporates SED template fitting constraints.
[REF_IMG]: PREML ArchitectureImpact
- Achieved SOTA accuracy, outperforming traditional methods.
- Satisfies two out of three LSST requirements up to z < 3.
- Created the PREML dataset (~400k samples), one of the largest public benchmarks.
Key Stats
~400k
Dataset Size
SOTA
Accuracy
z < 3
Redshift