In the development of the LHC as a precision-hadron collider,
the detector simulation has become a major limitation in terms of speed
and precision. In particular, reproducing the interactions of the
incident particles with the calorimeters and all the secondary
interactions is the slowest step of the simulation chain.
Without significant progress, first-principled simulations, based on
Geant4, will be the limiting
factor for all analyses at the high-luminosity upgrade of the LHC.
In this seminar, I will focus on the problem of building fast and
accurate
surrogate models for calorimeter shower simulations using cutting-edge
generative networks.
I will introduce the machine learning framework necessary to understand
generative networks and their evaluation. Then, I will discuss the
trade-off between generation speed and faithfulness by presenting two
approaches. First, a normalizing flow network is used to push the
sampling speed frontier.
Second, I will show how to maximize the quality of the showers
with a transformer-based conditional flow matching architecture at the
expense of a slower sampling speed.
