Particle-flow reconstruction with Transformer

¹ ETHZ, Zürich, Switzerland
² IP Paris, Paris, France
³ The University of Tokyo, Tokyo, Japan

Published online: 18 December 2024

Abstract

Transformers are one of the recent big achievements of machine learning, which enables realistic communication on natural language processing such as ChatGPT, as well as being applied to many other fields such as image processing. The basic concept of the Transformer is to learn relation between two objects by a self-attention mechanism. This structure is especially efficient with large input samples and large number of learnable parameters. We are studying this architecture applied to the particle-flow method, which reconstructs particles by clustering hits at highly-segmented calorimeters. Using datasets consisting of one or two initial photons, the network is asked to predict clusters one by one using hits from the calorimeters as input. Truth clusters information are provided at learning stage to compare with the decoder output. The best model reconstructed one photon events with a relative error on the energy of 5% and direction differing from the ground truth by 2.98 ◦. Moreover, the model achieved an accuracy of 99.6% when asked to separate one and two photons events.

This work was carried out in the framework of the ILD Concept Group