Elements: A new generation of samplers for astronomy and physics

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

Statistical inference is fundamental to scientific data analysis, and is often facilitated by the Bayesian methodology, which leverages posterior inference via Monte Carlo sampling. Traditionally, this has been a computational challenge, with the methodology often constituting the largest computational expense in many scientific fields, including but not limited to astronomy, cosmology, lattice quantum chromodynamics (QCD), and molecular dynamics. Fortunately, recent years have witnessed considerable academic progress in sampling methods, often resulting in significant computational reductions on synthetic examples. The primary objective of this project is to develop a software infrastructure that facilitates the use of these new and efficient samplers by a diverse range of scientists. The outcome of this research is anticipated to provide speedy and precise samplers that boast a user-friendly interface. These tools are not limited to applications in astronomy and physics but can be applied broadly across various scientific and engineering fields. Additionally, these innovative methodologies can be introduced into the curriculum of courses where statistical inference plays a key role, including courses on Data Science and Statistics for Science, thereby further enhancing the academic and practical understanding of these vital tools in the scientific community. The technical endeavor of this project involves developing software infrastructure for two cutting-edge samplers to facilitate Bayesian uncertainty quantification in various scientific fields. The first sampler, PocoMC, based on the Preconditioned Monte Carlo algorithm, utilizes Normalizing Flows and annealing for swift and accurate posterior analysis. The second sampler, the MicroCanonical Hamiltonian Monte Carlo, employs gradient-based methods effective in very high-dimensional scenarios where gradient-free methods fall short. Both samplers offer significant computational cost reductions compared to existing alternatives, presenting themselves as potential standard tools for scientists applying Bayesian methods. The project's scope encompasses the deployment of these new samplers as standalone packages and their integration into several widely-used Probabilistic Programming Languages, ultimately aiming to revolutionize statistical inference methods used in science and engineering. This award by the Office of Advanced Cyberinfrastructure is jointly supported by the Division of Astronomical Sciences and the Physics at the Information Frontier program in the Division of Physics within the Directorate for Mathematical and Physical Sciences. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.