Cosmological Parameter Estimation via Parameter-Efficient DenseNet and Tunable Loss Function

This presentation details the winning solution for Phase 1 of the FAIR Universe Weak Lensing ML Uncertainty Challenge. To address the challenge of estimating cosmological parameters (Ωm, S8) under simulation-model mismatch, this work proposes a highly parameter-efficient Deep Learning framework. The final winning model utilizes a DenseNet architecture (approx. 472k parameters), which prioritizes feature reuse over model depth. Empirical results demonstrated that this compact architecture provided a superior inductive bias for weak lensing maps, outperforming larger ResNet backbones in generalizing to unseen test data. A critical innovation driving this performance is a novel, tunable three-term loss function. By decoupling the standard Negative Log-Likelihood (NLL) into distinct weighted components—uncertainty penalty, weighted MSE, and pure MSE—the method enables precise control over the trade-off between point-estimate accuracy and uncertainty calibration. Supported by physics-informed noise injection, this approach highlights that architectural efficiency and optimized learning objectives are key to achieving trustworthy AI in cosmology.