Fast detector modeling using machine learning algorithms

Accurately and computationally rapidly modeling stochastic detector response for complex LHC experiments involving many particles from multiple interaction points, up to 200 interactions per proton-proton crossing in the HL-LHC requires the development of novel techniques. A study aimed at finding a fast transformation from truth level physics objects to reconstructed detector level physics objects is presented. This study used Delphes fast simulation based on an LHC-like detector geometry for inputs for machine learning (ML) algorithms, i.e. feed-forward regression neural networks, generative adversarial networks, and variational autoencoders. These ML transfer algorithms, with sufficient optimizations could have a wide range of applications to improve current detector simulations including: improving phenomenological studies by using a better detector representation, increasing the speed of creating event samples that more accurately resemble the output from Geant4-based detector simulation programs, or even speeding up fast simulations based on parametric description of LHC detector responses.