dN/dx Reconstruction with Deep Learning for High-Granularity TPCs

TL;DR

This work addresses the challenge of reconstructing $dN/dx$ in high-granularity TPCs for particle identification by introducing GraphPT, a graph-attention U-Net that treats detector hits as a point cloud. By incorporating graph neural networks with self-attention, GraphPT yields end-to-end $dN/dx$ reconstruction and outperforms the conventional truncated mean approach, achieving significant improvements in $K/\pi$ separation power across $5$–$20\,\mathrm{GeV}/c$ momentum. The study demonstrates that a dot-product attention variant provides the strongest PID performance, with competitive classification metrics (e.g., accuracy $\approx$0.707 and F1 $\approx$0.804) and notable gains over $dE/dx$ baselines. Looking ahead, the authors propose architectural and data diversification enhancements, computational optimizations, and beam-test validation to establish the practicality of $dN/dx$-based PID for next-generation detectors.

Abstract

Particle identification (PID) is essential for future particle physics experiments such as the Circular Electron-Positron Collider and the Future Circular Collider. A high-granularity Time Projection Chamber (TPC) not only provides precise tracking but also enables dN/dx measurements for PID. The dN/dx method estimates the number of primary ionization electrons, offering significant improvements in PID performance. However, accurate reconstruction remains a major challenge for this approach. In this paper, we introduce a deep learning model, the Graph Point Transformer (GraphPT), for dN/dx reconstruction. In our approach, TPC data are represented as point clouds. The network backbone adopts a U-Net architecture built upon graph neural networks, incorporating an attention mechanism for node aggregation specifically optimized for point cloud processing. The proposed GraphPT model surpasses the traditional truncated mean method in PID performance. In particular, the $K/π$ separation power improves by approximately 10% to 20% in the momentum interval from 5 to 20 GeV/c.

Figures (8)

• Figure 1: The is a cylindrical, gas-filled tracking detector whose axis is aligned with the nominal beam direction. The barrel houses a field cage assembly that generates a uniform electric drift field along the Z-axis. Two readout endplates provide mechanical support for the barrel, while a central cathode plane is located at its mid-plane.
• Figure 2: Working principle of a . 1) As a charged particle traverses the gas mixture, it generates a series of primary and secondary electrons. 2) These electrons drift toward the endcap under the influence of an electric field, undergoing diffusion during their transport. 3) Upon reaching the endcap, the electrons are amplified in the MicroMegas, producing a cluster whose spatial distribution has an effective width $w$.
• Figure 3: readout responses for a 2 cm track segment at the endcap. (a) Electrons before amplification, where the red circles represent primary electrons and the blue boxes represent secondary electrons. (b) Energy depositions on the readout pads, with the color bar indicating the deposited charge. (c) Pad labels on the readout pads, where the orange and white colors correspond to labels of positive and negative, respectively.
• Figure 4: Backbone structure of the . The network adopts a U-shaped encoder–decoder architecture with skip connections. Green boxes represent layers, yellow boxes indicate transformer layers, red boxes correspond to transition-down layers, and blue boxes correspond to transition-up layers. Within the encoder, nodes are randomly dropped from the graph, and the features are projected into a higher-dimensional space. The decoder performs the inverse operation. Finally, an additional layer is appended at the end of the U-Net.
• Figure 5: Optimization of the truncated mean method using 20 GeV/$c$ samples. (a) The $K/\pi$ separation power as a function of the fraction $\alpha$, with the number of pad rows fixed to two. (b) The $K/\pi$ separation power as a function of the number of pad rows to be combined, with the fraction fixed at $\alpha = 0.65$. The green solid line represents the d$N$/d$x$ method, while the orange dashed line represents the d$E$/d$x$ method. The optimal values are $\alpha = 0.65$ and two pad rows, respectively.