Vision transformers outperform CNNs in segmenting cosmic proto-halos

Researchers have developed deep learning models, specifically a U-Net transformer and a V-Net-based CNN, to segment proto-halos in the early universe's density field. The transformer-based network demonstrated superior performance, achieving sub-percent error in segmented mass per halo class and outperforming the traditional extsc{pinocchio} model, particularly for low-mass halos and boundary reconstruction. The study also examined the influence of input features like density and tidal shear, and used Grad-CAM to visualize the CNN's internal workings. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Demonstrates advanced deep learning techniques for complex scientific simulations, potentially improving cosmological modeling.

RANK_REASON Academic paper detailing a new application of vision transformers and CNNs for astrophysical simulations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• V-Net
• U-Net
• pinocchio
• Grad-CAM
• Toka Alokda

• paper
• other