Deep(er) Reconstruction of Imaging Cherenkov Detectors with Swin Transformers and Normalizing Flow Models

TL;DR

This work tackles the challenge of fast, accurate particle identification and realistic hit-pattern simulation in imaging Cherenkov detectors. It introduces Deep(er)RICH, which fuses Swin Transformer-based image processing with Normalizing Flows that model conditional hit-density PDFs as a function of track kinematics, enabling both near real-time PID and high-fidelity, data-driven fast simulations for the GlueX DIRC. The approach outperforms traditional LUT-based geometry across the full phase space, achieving about 9 µs per particle for inference and 0.5 µs per hit for generation on GPU hardware, while validating that fast-sim results are statistically indistinguishable from Geant4-trained data. This combination of high-accuracy PID and fast, portable simulation holds significant potential for real-time triggering, online calibration, and extended applicability to future imaging Cherenkov systems such as ePIC’s DIRC/dual-RICH configurations. All told, Deep(er)RICH advances practical detector understanding by learning directly from data and enabling scalable, physics-aware simulations for complex topologies.

Abstract

Imaging Cherenkov detectors are crucial for particle identification (PID) in nuclear and particle physics experiments. Fast reconstruction algorithms are essential for near real-time alignment, calibration, data quality control, and efficient analysis. At the future Electron-Ion Collider (EIC), the ePIC detector will feature a dual Ring Imaging Cherenkov (dual-RICH) detector in the hadron direction, a Detector of Internally Reflected Cherenkov (DIRC) in the barrel, and a proximity focus RICH in the electron direction. This paper focuses on the DIRC detector, which presents complex hit patterns and is also used for PID of pions and kaons in the GlueX experiment at JLab. We present Deep(er)RICH, an extension of the seminal DeepRICH work, offering improved and faster PID compared to traditional methods and, for the first time, fast and accurate simulation. This advancement addresses a major bottleneck in Cherenkov detector simulations involving photon tracking through complex optical elements. Our results leverage advancements in Vision Transformers, specifically hierarchical Swin Transformer and normalizing flows. These methods enable direct learning from real data and the reconstruction of complex topologies. We conclude by discussing the implications and future extensions of this work, which can offer capabilities for PID for multiple cutting-edge experiments like the future EIC.

Figures (7)

• Figure 1: Schematic of GlueX DIRC geometry: A charged particle traverses the fused silica bar (step 1), generating Cherenkov light (step 2). This light undergoes internal reflection, reaching the optical box. Through a series of mirrors, the light is directed to an array of PMTs (step 3) for detection. The resulting hit pattern, which depends on the particle's kinematics, is illustrated in the final step, showing the accumulated hit patterns from multiple tracks with the same kinematics.
• Figure 2: Optical box output: Individual tracks leave sparse hit patterns (red points) integrated over time on the DIRC readout plane, proving to be a challenge for convolutional-based networks to deal with. The denser hit pattern is obtained by accumulating multiple tracks with same kinematics. White zones are locations at which PMTs are not installed due to low accumulation of hits.
• Figure 3: Architecture flow chart: Images of the optical box are processed through four consecutive Swin encoder blocks, producing feature maps at different resolutions. These outputs are fed in parallel to a Convolution Neural Network for recombination and downsampling prior to a flattening operation. The flattened vector is concatenated to the track kinematics and processed by a simple Deep Neural Network producing a binary label.
• Figure 4: Flow chart of Delta-Loglikelihood with two Normalizing Flows: Individual tracks are represented by matrices of individual Cherenkov photons, conditional on the kinematic parameters. We compute the likelihood of each Cherenkov photon in the base distribution of the normalizing flow for each hypothesis $\pi / K$, such that the total likelihood is the summed contribution of individual hits. The summed quantities are then used to form a DLL on a track-by-track basis.
• Figure 5: Particle gun performance: Pion rejection as a function of kaon efficiency for the Swin architecture, Normalizing Flow method, compared to the standard geometrical method, integrated over the entire phase space (left). The Area Under the Curve (AUC) is indicated within the legend. AUC as a function of track momentum (right). Uncertainty is represented as the 95% quantiles obtained through bootstrapping. The counts of pions and kaons for testing is also reported.