Gsoc2026 evaluations

[abhiram123467]

GSoC 2026 Evaluation Submission

Applicant: Abhi Ramg
Project: ML4Sci — DeepLense

This Pull Request contains my evaluation notebook submissions for the GSoC 2026 DeepLense project. The repository includes three separate tasks demonstrating proficiency in Foundation Models, Diffusion Models, and Super-Resolution architectures for gravitational lensing.

🔭 Task Breakdown & Results

1. DEEPLENSE4: Foundation Model for Gravitational Lensing

• Methodology: Built a Vision Transformer (ViT) Masked Autoencoder (MAE) inspired by Lens-JEPA.
• Pre-training: Self-supervised MAE pre-training on no_sub images with a 75% masking ratio to learn underlying structural physics.
• Task IX.A (Classification): Fine-tuned the encoder for 3-class dark matter classification (CDM, Axion, No Substructure) achieving robust ROC-AUC scores.
• Task IX.B (Super-Resolution): Reused the frozen encoder backbone and fine-tuned a custom decoder for 4x super-resolution.

2. DEEPLENSE8: Physics-Informed Diffusion Models

• Methodology: Implemented a Denoising Diffusion Probabilistic Model (DDPM) with a custom U-Net backbone (incorporating Sinusoidal Time Embeddings and Attention blocks).
• Training: Trained on 10,000 strong lensing images over 1000 diffusion timesteps.
• Evaluation: Generated highly realistic lensing images from pure noise, evaluated using Frechet Inception Distance (FID) to quantify domain-specific realism.

3. DEEPLENSE9: Unsupervised Super-Resolution

• Methodology: Developed a Residual Channel Attention Network (RCAN-lite) to reconstruct high-resolution lensing features from degraded inputs.
• Task VI.A (Simulated): Trained on simulated HR/LR pairs, optimizing a combined L1 + Gradient loss function to preserve critical arc sharpness.
• Task VI.B (Real HSC/HST): Applied transfer learning by fine-tuning the Task A model on 300 real telescope image pairs to handle physical noise distributions (Domain Adaptation).
• Metrics: Evaluated against Bicubic baselines using MSE, SSIM, and PSNR.

📁 Attached Files

• DEEPLENSE4_Foundation_Model_FINAL.ipynb
• DEEPLENSE8_DDPM_FINAL.ipynb
• DEEPLENSE9_SuperResolution.ipynb
• Related output visualizations and metric dashboards.