11:00 am–12:00 pm ERC 401
Host: Leah Jenks
Michael Toomey ( MIT) "Normalizing Flows for Cosmological Data Analysis: Learning Physically Motivated Priors in Early Dark Energy and Full-Shape Analyses"
Machine learning (ML) has been rapidly implemented for a variety of analyses in the era of precision cosmology. To date, applications have generally leveraged large quantities of readily available data, or even simulations, as a means to extract insights about the underlying physics. One area that has not benefited nearly as much from these advances is cosmological parameter inference and model comparison. The limited adoption of machine learning in this context is understandable due to a serious drawback: it is by no means transparent how ML algorithms converge during training. By comparison, parameter inference and model comparisons for ΛCDM and its extensions, using posterior sampling from Markov Chain Monte Carlo, are well-posed given concrete theoretical predictions and cosmological likelihoods. However, in such a pipeline, the choice of priors represents at least a somewhat ambiguous choice, especially for new and previously unconstrained parameters. Typically, this situation is addressed by adopting broad, uninformative priors on parameters. This approach has the downside that it may ignore existing knowledge of the underlying physics in a given model, especially when a certain representation or parameterization is not possible to integrate into the analysis pipeline directly. To bridge this gap, this talk introduces neural density estimators, specifically normalizing flows, as a useful tool for constructing physically informed priors on cosmological parameters. This technique allows for the training of normalizing flows in domains where the underlying physics is better understood, which can then be seamlessly integrated into a standard analysis pipeline with only a few lines of new code. One of the benefits of this approach is the potential to both improve model constraining power and significantly speed up the analysis process. As a concrete demonstration, this technique will be applied to two cases: first, using normalizing flows to learn a prior for parameters in the early dark energy model, informed by theory, demonstrating significant improvements in constraints and a speed increase of the analysis by more than two orders of magnitude. In the second case, we train a normalizing flow on simulated galaxy catalogs to learn a prior for EFT bias parameters. We then use our learned prior to improve constraints on single-field inflation by nearly a factor of two in a full-shape analysis with BOSS. This work not only highlights the versatility of normalizing flows in cosmological data analysis but also sets the stage for future advancements in applying machine learning techniques to better bridge the gap between theory and data analysis.
