SAIP2026

Deploying physics-informed neural networks for anomaly detection in particle physics

9 Jul 2026, 12:00

20m

Lecture Hall DL3 (University of the Western Cape)

Oral Presentation Track F - Applied Physics Applied Physics

Speaker

Dr Makhamisa Senekane (University of Johannesburg)

Description

The Standard Model (SM) of particle physics provides a mathematical description of the constituents of matter, together with their interactions. The deviation from the SM might lead to the outliers that herald the new physics. Therefore, the use of anomaly detection models plays a crucial role in the discovery of the new physics; thereby facilitating the move beyond the Standard Model. For anomaly detection, deep learning (DL) models; especially those based on the autoencoder architecture, are typically used. However, these data-driven autoencoders might produce results that do not correspond to the physical reality. Therefore, in order to address this challenge (of lack of correspondence to physical reality, a more realistic neural networks architecture is required. One such architecture is the physics-informed neural networks (PINN), which provides the physical reality check by ensuring that the output of the neural networks model is consistent with the laws of physics. In this paper, we report the use of PINN autoencoder model for anomaly detection in particle physics. Finally, we demonstrate that the fusion of the laws of physics with the data-driven DL leads to improved results in the detection of anomalies in particle physics.

Presentation materials

169_Makhamisa_Senekane.pdf