ParaFlow: fast calorimeter simulations parameterized in upstream material configurations

We study whether machine-learning models for fast calorimeter simulations can learn meaningful representations of calorimeter signatures that account for variations in the full particle detector's configuration. This may open new opportunities in high-energy physics measurements, for example in the assessment of systematic uncertainties that are related to the detector geometry, in the inference of properties of the detector configuration, or in the automated design of experiments. As a concrete example, we parameterize normalizing-flow-based simulations in configurations of the material upstream of a toy calorimeter. We call this model ParaFlow, which is trained to interpolate between different material budgets and positions, as simulated with Geant4. We study ParaFlow's performance in terms of photon shower shapes that are directly influenced by the properties of the upstream material, in which photons can convert to an electron-positron pair. In general, we find that ParaFlow is able to reproduce the dependence of the shower shapes on the material properties at the few-percent level with larger differences only in the tails of the distributions.