Experiments with tiny particles of large mass—such as compressed baryonic matter (CBM)—are an important pillar of the Facility for Antiproton and Ion Research (FAIR) in Darmstadt. The CBM experiment at the FAIR accelerator will record up to ten million heavy ion collisions per second. Under these conditions, an online trigger is required to select events that may contain signatures of quark-gluon plasma. In his dissertation, supervised by FIAS Fellow Ivan Kisel, Artemiy Belousov investigated how neural-network approaches can support this task.

In his work, Belousov presents deep neural networks trained to recognize Quark-Gluon-Plasma formation directly from simulated data. The models are integrated into ANN4FLES, a  high-performance framework used in the First Level Event Selection (FLES) system. They achieve high identification accuracy and remain stable when evaluated with different types of input, including reconstructed detector outputs.

The study also applies game theory techniques to explain the output of any machine learning model: SHAP (SHapley Additive exPlanations) allows the contribution of individual features to the network's decisions to be analyzed. This allows the behaviour of the trigger to be interpreted in terms of physical quantities relevant to the formation of quark-gluon plasma.

The results demonstrate that neural-network methods can provide fast and reliable event selection for CBM and support the experiment’s physics goals.

Belousov successfully presented his data on November 26 and afterwards celebrated his doctorate. During his studies he was also very engaged in social activities at FIAS. Belousov is now continuing his research as a postdoc in the group of FIAS Senior Fellow Volker Lindenstruth.