COLLOQUIUM "Learning from Simulations: AI and the Future of Inference in Gravitational-Wave and Cosmological Data" by Christoph Weniger (UvA) as part of the National Seminar Theoretical High Energy Physics

Friday 14 Nov 2025, 11:15 → 12:15 Europe/Amsterdam

 

 

Modern gravitational-wave and cosmological analyses increasingly rely on simulations rather than analytic models. Traditional inference methods struggle with overlapping signals, non-Gaussian noise, and the enormous parameter spaces of Stage IV surveys and future observatories. Simulation-based inference (SBI) offers a new approach: by training neural networks on synthetic data, we can learn to perform Bayesian inference directly from simulations. I will illustrate this idea through examples from gravitational-wave parameter estimation — including stochastic backgrounds and extreme-mass-ratio inspirals — and from field-level cosmological inference. I will also discuss distributed, dynamically adaptive SBI frameworks that enable parallelized and online analysis at scale. The talk will conclude with a reflection on how these developments challenge our notions of transparency, interpretability, and trustworthiness — and how we can ensure that AI-driven inference remains a reliable and explainable tool for future experiments such as LISA.