Predicting the Exclusive Diffractive Electron-Ion Cross Section at small $x$ with Machine Learning in Sar$t$re

Abstract

The event generator Sar$t$re has been used extensively for simulations of electron-ion collisions in preparation for the Electron-Ion Collider (EIC). Sar$t$re simulates exclusive diffraction in $e$A collisions, in principle for any nuclear species and exclusive final state, usually a vector meson. The coherent and incoherent cross sections for each process are calculated in the colour dipole model for small $x$ from the first and second moments of the respective amplitude, averaged over initial state spatial configurations. Taking these averages is a very CPU demanding task. In order to function as an efficient event generator, these amplitude moments are saved into lookup tables which are used as input for the event generation, making the latter a very fast process. However, there are many recent and ongoing developments of the dipole models underlying the calculations, both in terms of fits of the model parameters to new data as well as new parametrisations of the dipole or proton geometries. Therefore, it is desirable to have a more flexible method for producing the lookup tables. Here, we propose a method using neural networks which can reduce the table production time by 90% while retaining the same precision in the resulting cross sections.