Decoding the Universe with AI: A Decade of Progress and the Road Ahead

Weak gravitational lensing (i.e. small deformation of the shapes of faraway galaxies when their light is deflected by massive objects along the line of sight) has become one of the most powerful probes of the dark Universe. Yet fully exploiting the information content of this lensing effect remains one of the field’s greatest statistical and computational challenges. Over the past decade, the convergence of cosmology, modern machine learning, and simulation-based inference (SBI) has reshaped our ability to extract information from the data and learn about our Universe. In this keynote, I will provide a domain expert perspective on the evolution of AI-driven cosmological inference. I will discuss the scientific opportunities unlocked by these tools, their current limitations driven by simulation fidelity and domain shift, and the methodological advances needed to ensure trustworthy cosmological inference for modern large-scale cosmological surveys. Finally, I will reflect on what we learned from the FAIR Universe NeurIPS Challenge and how these community efforts can accelerate progress toward reliable, interpretable, and simulation-efficient methods for the next generation of cosmology.