11 Detector Simulation
505
Recently ATLAS has produced an integrated simulation framework ISF [98],
which includes detailed simulation and per-subdetector fast simulations. This
enables the use of common modules for all elements of the simulation. A hybrid
simulation mode is part of its design that selects conical regions around the most
energetic primary particles for detailed simulation, and uses fast simulation models
for the remaining parts of the detector. This capability has not yet been used in
physics analyses.
11.4 Machine Learning for Fast Simulation
A completely new approach to parameterized simulation has emerged recently,
exploring the potential of machine learning. One research avenue attempts to
generate patterns of energy deposition that reproduced the distributions including
fluctuations of key physical observables.
This has been the topic of interest and recent investigations using the generativeadversarial network (GAN) approach to the fast generation of patterns of hits. A
first demonstration in a three layer LAr sampling calorimeter [99] was tested over
an energy range from 1 to 100 GeV and a single particle direction. A potential
speedup of O(100)–O(10,000) was demonstrated. Most physical observables of
interest for the classification and calibration of tracks were well reproduced, but
a few showed clear differences. This promising avenue will clearly be an area of
significant research in the next years.
11.4.1 Accuracy of Simulation
The accuracy of the simulation is determined not only by the artificial differences
or defects introduced by such simplifications, but also by intrinsic factors. These
factors include the accuracy of the cross-sections for particular interactions and
the capability of the physical models. These can be explored by comparing with
experimental data.
Criticisms of detector simulation focus on key limitations and question the
predictive power of hadronic interaction modeling for use in designing and tuning
hadronic calorimeters [100].
11.4.2 Signal Generation
In order to model the signal produced in a gaseous detector, all processes that
contribute to the generation of charge and its collection in the cathode must be
simulated. The detailed simulation of a small number of events modeled in full detail
is used to understand the characteristics of a detector. For large-scale simulation, a
505
Recently ATLAS has produced an integrated simulation framework ISF [98],
which includes detailed simulation and per-subdetector fast simulations. This
enables the use of common modules for all elements of the simulation. A hybrid
simulation mode is part of its design that selects conical regions around the most
energetic primary particles for detailed simulation, and uses fast simulation models
for the remaining parts of the detector. This capability has not yet been used in
physics analyses.
11.4 Machine Learning for Fast Simulation
A completely new approach to parameterized simulation has emerged recently,
exploring the potential of machine learning. One research avenue attempts to
generate patterns of energy deposition that reproduced the distributions including
fluctuations of key physical observables.
This has been the topic of interest and recent investigations using the generativeadversarial network (GAN) approach to the fast generation of patterns of hits. A
first demonstration in a three layer LAr sampling calorimeter [99] was tested over
an energy range from 1 to 100 GeV and a single particle direction. A potential
speedup of O(100)–O(10,000) was demonstrated. Most physical observables of
interest for the classification and calibration of tracks were well reproduced, but
a few showed clear differences. This promising avenue will clearly be an area of
significant research in the next years.
11.4.1 Accuracy of Simulation
The accuracy of the simulation is determined not only by the artificial differences
or defects introduced by such simplifications, but also by intrinsic factors. These
factors include the accuracy of the cross-sections for particular interactions and
the capability of the physical models. These can be explored by comparing with
experimental data.
Criticisms of detector simulation focus on key limitations and question the
predictive power of hadronic interaction modeling for use in designing and tuning
hadronic calorimeters [100].
11.4.2 Signal Generation
In order to model the signal produced in a gaseous detector, all processes that
contribute to the generation of charge and its collection in the cathode must be
simulated. The detailed simulation of a small number of events modeled in full detail
is used to understand the characteristics of a detector. For large-scale simulation, a