Transformer-based machine learning using low-level calorimeter signals for collimated photon identification at collider experiments

Abstract

Electromagnetic calorimeters provide essential information for reconstructing and selecting both Standard Model (SM) and potential beyond the SM physics events at high-energy particle colliders. The fine-grained segmentation of modern calorimeters captures rich information about the internal structure of particle showers, much of which is discarded by conventional high-level reconstruction methods. In this work, we leverage calorimeter cell-level information to classify highly collimated diphoton signatures, arising from the decay of light axion-like particles, from isolated single-photon showers. We systematically compare a range of machine learning architectures, spanning high-level, shower shape variable-based approaches and direct cell-level methods. Cell-level machine learning shows significantly superior classification ability, with a Transformer in particular representing the best performance among six different architectures studied, and an MLP Mixer representing a resource-constrained alternative for potential real-time, trigger-level applications. Beyond classification, the Transformer model developed enables direct invariant mass regression from calorimeter cells, improving the characterization of light resonances and providing an additional handle in reducing the $\pi^0$ and $\eta$ fake photon backgrounds. These results demonstrate that cell-level machine learning methods can extend calorimeter-based particle identification and performance well beyond the capabilities of current conventional techniques.