Cosmological Parameter Estimation with a Denoising U-Net and Patch-Based CNNs

This presentation describes a multi-step Deep Learning framework developed for Phase 1 of the FAIR Universe Weak Lensing ML Uncertainty Challenge. The approach estimates the cosmological parameters (Ωₘ, S₈) from noisy convergence maps using two main stages. The pipeline first applies a denoising U-Net to reduce pixel-level noise and produce cleaner inputs. The denoised maps are then processed by patch-based CNNs with an attention mechanism, which extract both local and global patterns relevant for cosmological inference. To further improve robustness and reduce variance, the final estimates are obtained by ensembling multiple CNN backbones. Overall, the method combines denoising, patch-based CNNs with attention, and ensembling to deliver stable and accurate parameter estimates.