README.md

Gravitational Lensing

Dark matter is a hypothetical form of matter, which does not interact with electromagnetic radiation. It does not reflect or emit light and is not directly observable by the human eye. However, its existence can be inferred by observing gravitational effects on visible matter, such as stars and galaxies. Gravitational lensing causes light to bend in the presence of a strong gravitational field. Due to this bending of light, distant objects may appear to be distorted or magnified. The study of these distorted shapes can aid researchers in identifying the distribution and location of dark matter. By analysing a variety of different images, it is possible to deduce the distribution of dark matter. Furthermore, by measuring the distortion geometry, the mass of the surrounding cluster of dark matter can be determined. This project performs 3 fundamental tasks related to dark matter and gravitational lensing:

• Binary Substructure Classification
• Dark Matter Halo Mass Prediction
• Multiclass Substructure Classification

Binary Substructure Classification

Deep Learning Model	Epochs	Batch Size	Learning Rate	ROC AUC
ViT_Base_Patch_16_224	20	64	0.0001	0.99800

Results for ViT_Base_Patch_16_224 (20 Epochs):

Dark Matter Halo Mass Prediction

Models Used:

Deep Learning Model	Epochs	Batch Size	Learning Rate	MSE
EfficientNetB4	10	128	0.0005	0.0002007
ConvNeXtBase	25	128	5e-05	0.0002763
InceptionResNetV2	20	128	5e-05	0.0002618

Results for EfficientNetB4 (10 Epochs):

Results for ConvNeXtBase (25 Epochs):

Results for InceptionResNetV2 (20 Epochs):

Multiclass Substructure Classification

Models Used:

Deep Learning Model	Epochs	Batch Size	Learning Rate	ROC AUC (OvO)	ROC AUC (OvR)
DenseNet161	15	64	0.0001	0.98	0.98
MobileVitV2_150_384_in22ft1k	15	32	0.0001	0.95	0.95
DenseNet201	15	64	0.0001	0.97	0.97
Ensemble_DenseNet161_DenseNet201	10	32	0.0001	0.98	0.98

Results for DenseNet161 (15 Epochs):

Results for MobileVitV2_150_384_in22ft1k (15 Epochs):

Results for DenseNet201 (15 Epochs):

Results for Ensemble Model (DenseNet161 & DenseNet201 for 10 Epochs):

Usage

Clone the repository

https://github.com/rprkh/Gravitational-Lensing.git

Navigate to the root directory of the project

cd Gravitational-Lensing

Install the requirements

pip install -r requirements.txt

Run the following command

mkdir "models\binary_substructure_classification" "models\dark_matter_halo_mass_prediction" "models\multiclass_substructure_classification"

Download the trained models from the following Google Drive link: https://drive.google.com/drive/folders/1NAeesQqyHlF6mu7Uv8sXiZx5eaRC3JBy?usp=sharing

Add these models to their respective folders within the models directory of the project

Execucte the following command

streamlit run streamlit_app.py

The streamlit application should start on http://localhost:8501/