Parts | Days | Selection | Search | Updates | Downloads | Help
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 10: AI Topical Day – Computing II (joint session HK/AKPIK)
AKPIK 10.1: Talk
Thursday, March 23, 2023, 14:00–14:15, HSZ/0103
Exploiting Differentiable Programming for the End-to-end Optimization of Detectors — The MODE Collaboration1 and •Anastasios Belias2 — 1mode-collaboration.github.io — 2GSI Helmholtzzentrum für Schwerionenforschung GmbH, Darmstadt, Germany
Machine-learning Optimized Design of Experiments, the MODE Collaboration, targets the end-to-end optimization of experimental apparatus, by using techniques developed in modern computer science to fully explore the multi-dimensional space of experiment design solutions. Differentiable Programming is employed to create models of detectors that include stochastic data-generation processes, the full modeling of the reconstruction and inference procedures, and a suitably defined objective function, along with the cost of any given detector configuration, geometry and materials.
The MODE Collaboration considers the end-to-end optimization challenges in its generality, providing software architectures for machine learning to explore experiment design strategies, information on the relative merit of different configurations, with the potential to identify and investigate novel, possibly revolutionary solutions. In this contribution we present use cases, and highlight the potential for on-going and future experiment design studies in fundamental physics research.