Jan Kieseler (CERN)
Application and development of advanced deep neural networks for high-granularity calorimeters
Abstract: High-granularity calorimeters, in particular those
providing also fine-grained lateral resolution, offer unique
possibilities to resolve the development of individual
particle showers, separate electromagnetic and hadronic
contributions on a shower-by-shower basis and perform
calorimeter-driven identification tasks. Furthermore, they can
be operated as tracking detectors makes it possible to follow
the trajectories e.g. of muons passing through and connect
individual parts of hadronic showers, provided the energy
threshold is below MIP level. Analytic solutions based on
simplified models and few tuneable parameters might not be
able to harness the full potential in particular for hadronic
showers. Machine learning techniques, however, have been very
successful in pattern recognition tasks for images or - lately
- point clouds, where algorithmic solutions tend to be overly
complex and - at the same time - by far do not reach the same
performance.
The seminar will cover applications of neural network
architectures to shower segmentation, particle identification
and energy regression in high granularity calorimeters. The
detector geometries and physics considerations call for
customised solutions, therefore the focus will be on dedicated
developments, such as special graph neural networks, and how
they can be applied physics problems. This demand for new
solutions triggers machine-learning developments beyond
calorimeters with impact on other areas of reconstruction such
as tracking and global event reconstruction, e.g. particle
flow, but also beyond particle physics. Please note that a
Webcast retransmission will be available for this Seminar.