Efficient Forward-Mode Algorithmic Derivatives of Geant4

TL;DR

This paper presents a novel application of forward-mode algorithmic differentiation to Geant4, enabling gradient computations for Monte Carlo simulations, which facilitates optimization in high-energy physics and related fields.

Contribution

It introduces a differentiated version of Geant4 using operator-overloading AD, allowing mean pathwise derivatives of outputs to be computed with respect to inputs.

Findings

Successfully applied AD to Geant4 for derivative computation

Demonstrated potential for gradient-based optimization in physics simulations

Initial analysis shows promising results for future refinement

Abstract

We have applied an operator-overloading forward-mode algorithmic differentiation tool to the Monte-Carlo particle simulation toolkit Geant4. Our differentiated version of Geant4 allows computing mean pathwise derivatives of user-defined outputs of Geant4 applications with respect to user-defined inputs. This constitutes a major step towards enabling gradient-based optimization techniques in high-energy physics, as well as other application domains of Geant4. This is a preliminary report on the technical aspects of applying operator-overloading AD to Geant4, as well as a first analysis of some results obtained by our differentiated Geant4 prototype. We plan to follow up with a more refined analysis.