CaloFlow II: Even Faster and Still Accurate Generation of Calorimeter Showers with Normalizing Flows

The enormously successful physics program at the LHC relies heavily on the availability of copious amounts of highly accurate simulated data. However, the use of Geant4 [1-3] for full detector simulations is a major computational bottleneck and severely limits the analysis capabilities of the LHC. This is forecast to worsen significantly with future LHC upgrades and the HL-LHC [4-8]. Recently, deep generative modeling has demonstrated great potential to speed up the most computationally expensive part of detector simulations, namely calorimeter showers [8-19]. By fitting the generative model to Geant4 shower images, the generative model learns (often implicitly) the underlying distribution that the Geant4 showers are drawn from and can then sample from it quickly. Most of the current approaches [8-18] are based on GAN or VAE architectures. Very recently, in [19], we proposed a fresh alternative, dubbed CaloFlow, based on normalizing flows (for recent reviews and original references, see e.g.