Differentiating a HEP Analysis Pipeline within the Scikit-HEP Software Ecosystem

TL;DR

This paper introduces a differentiable analysis pipeline for high-energy physics that leverages JAX within the Scikit-HEP ecosystem, enabling end-to-end parameter tuning and improved sensitivity in particle searches.

Contribution

It presents the first differentiable HEP analysis pipeline combining Scikit-HEP tools with JAX, demonstrating enhanced optimization capabilities and statistical significance in a real-world collider data analysis.

Findings

Significant improvement in expected statistical significance.

Successful end-to-end parameter tuning using gradient-based methods.

Practical strategies for integrating differentiable techniques in HEP workflows.

Abstract

A first differentiable analysis pipeline is presented for an example high-energy physics (HEP) use case with publicly available collision data from the Compact Muon Solenoid detector at the Large Hadron Collider. The pipeline combines tools from the Scikit-HEP ecosystem with JAX. The study is based on an existing search for a hypothetical particle, the $Z^{\prime}$ boson, and uses a realistic, yet simplified, statistical model. The gradient-based optimization techniques employed in this work can advance HEP workflows by enabling end-to-end tuning of analysis parameters, improving both computational scalability and overall sensitivity. The challenges of adopting such techniques in HEP workflows are highlighted, along with practical mitigation to those challenges. This framework results in a significant improvement in expected statistical significance compared to a baseline analysis by fine-tuning $\mathcal{O}(10^3)$ parameters in the pipeline. Perspectives on future applications and recommendations for broader engagement with differentiable techniques in the field are also outlined.