Toward the end-to-end optimization of the design of experiments

Name: Toward the end-to-end optimization of the design of experiments
Start: 2021-11-02T11:00:00+01:00
End: 2021-11-02T12:00:00+01:00
Location: No location set

by Tommaso Dorigo (INFN Padova)

Tuesday 2 Nov 2021, 11:00 → 12:00 Europe/Stockholm

Description

In the past few decades particle physics has made giant leaps by studying matter at the shortest distance scales with detectors built on a few paradigms (track-first-destroy-later, redundancy, cross-verification and calibration) which worked very well, but which today look increasingly misaligned with the progress of information-extraction procedures, as the construction choices of complex apparatus live in hundreds-dimensional parameter spaces which until recently humans have been unable to probe systematically.

The result of not investigating the full space of solutions to particle detection, pattern recognition, and inference extraction is a huge potential loss in performance. Our saviour may be differentiable programming, which allows the creation of continuous models of all the ingredients of measurements in fundamental physics, as well as, crucially, a carefully constructed objective function. Backpropagation through a differentiable pipeline of all the elements of the problem may allow to probe the design space and realign design and goals of experiments that base their operation on the interaction of radiation with matter, besides finding entirely new ways to solve our detection problems.

In this seminar I will illustrate how we can set out to create the interfaces to solve our difficult optimization problems, with great prospects for future endeavours in HEP, astro-HEP, nuclear and neutrino physics when these allow for time and resources to carry out such studies.