Galin Bistrev

CERN Summer student 2025

Education: Nuclear and particle physics , Sofia University , St. Kliment Ohridski , Bulgaria

Ongoing project: Implementing AD in CMS Combine
This project aims to introduce automatic differentiation techniques into the CMS Combine tool. Combine is the primary statistical analysis tool used in CMS and is based on the RooFit software. Recently RooFit has started using automatic differentiation (AD) technology to aid minimization algorithms. Computationally cheap gradients produced by AD can improve performance of the larger workflows in Combine by more than 5x as demonstrated at ICHEP. The way how RooFit implements AD also benefits the shareability of likelihoods. To get the likelihood gradient, RooFit transforms the likelihood from its internal representation to simple standalone C++ code, which is then used to generate the gradient code. This can help collaborators without RooFit and Combine expertise to work with likelihoods from Combine at a lower level. Adopting AD in Combine requires further collaboration between the parties. This project aims to bring this AD support into Combine. Aspects include reducing some of the code in Combine in favor of more standard RooFit primitives; Implementing AD support for Combine-specific RooFit primitives; and writing benchmarks to showcase the benefits of RooFit AD in Combine. The student selected for this project will work with several high-energy physics codes as well as the LLVM compiler ecosystem.

Project Proposal: URL

Mentors: Vasil Vassilev,David Lange,Jonas Rembser

Galin Bistrev

Presentations