Cluster Counting (dN/dx) in Particle Detectors

• Cluster Counting (dN/dx) is defined as a particle identification method that measures primary ionization clusters per unit length, leveraging Poisson-like statistics.
• It improves resolution by targeting microphysical ionization features in helium-based drift chambers and TPCs, leading to enhanced hadron separation.
• Modern reconstruction strategies, including derivative-based algorithms and machine learning techniques, effectively address challenges like cluster overlap and inefficiencies.

Cluster counting, commonly denoted $dN/dx$ or $dN_{\mathrm{cl}}/dx$, is a gaseous-detector particle-identification technique in which the measured ionization observable is the number of primary ionization clusters produced per unit track length rather than the total deposited energy $dE/dx$. The method targets the discrete primary-ionization process directly, and is therefore motivated by the approximately Poissonian statistics of cluster production, in contrast to the Landau-like fluctuations and $\delta$-ray sensitivity that dominate charge-integration observables. Across studies of helium-based drift chambers and high-granularity time projection chambers, $dN/dx$ is consistently treated as a route to narrower PID response distributions and stronger hadron separation than conventional $dE/dx$, provided that the detector and readout preserve enough temporal or spatial structure to resolve individual ionization features (Caputo et al., 2022, Cuna et al., 2021).

1. Physical basis of the observable

In the conventional $dE/dx$ approach, a track is characterized by the total energy deposited per unit length, and a truncated mean is usually applied to suppress the long high-energy tail generated by rare, highly ionizing deposits and energetic secondary electrons. Cluster counting replaces that analog observable with a counting observable: if $N$ primary clusters are produced along a path length $x$, then the PID variable is approximately

$\frac{dN}{dx}\approx \frac{N}{x}.$

For a gas volume of length $dN_{\mathrm{cl}}/dx$0, the expected number of clusters is written as

$dN_{\mathrm{cl}}/dx$1

This formulation appears throughout the drift-chamber literature as the basic definition of the method (Cuna et al., 2021, Elmetenawee et al., 26 Sep 2025).

The statistical argument for $dN_{\mathrm{cl}}/dx$2 is central. Primary-cluster formation is treated as a counting process that is closer to Poisson behavior than the broad charge-deposition spectrum used in $dN_{\mathrm{cl}}/dx$3. Several studies therefore emphasize that cluster counting is less sensitive to cluster-size fluctuations, gas-gain variations, energetic secondaries, and Landau tails, and that its resolution should improve with track length in a Poisson-like manner. In beam-test analyses of helium-based drift chambers, the observed cluster-count distribution is described as Gaussian with width almost equal to $dN_{\mathrm{cl}}/dx$4, and dedicated resolution studies find $dN_{\mathrm{cl}}/dx$5 scaling for $dN_{\mathrm{cl}}/dx$6, compared with $dN_{\mathrm{cl}}/dx$7 for $dN_{\mathrm{cl}}/dx$8 (Caputo et al., 2022, Elmetenawee et al., 26 Sep 2025).

The PID discrimination is usually expressed through a separation-power metric. One formulation used for CEPC drift-chamber studies is

$dN_{\mathrm{cl}}/dx$9

while high-granularity TPC studies also use

$dE/dx$0

Both definitions quantify how far apart the species-dependent $dE/dx$1 response distributions are relative to their widths, and both are used to compare cluster counting with $dE/dx$2 or truncated-mean baselines (Tian et al., 2024, Zhao et al., 12 Oct 2025).

2. Detector media, granularity, and readout conditions

The method is especially associated with helium-based gases, because low-density mixtures help preserve the separability of the underlying primary-ionization structure. In the IDEA drift chamber, the helium-based gas is characterized by a low drift velocity of about $dE/dx$3, a comparatively large cluster time separation of about $dE/dx$4 in $dE/dx$5 He, a small mean cluster size of $dE/dx$6, and low single-electron diffusion. These properties are identified as favorable because cluster counting requires primary-ionization “blobs” to remain distinguishable at the anode-signal level (Caputo et al., 2022).

The detector geometry and sampling scheme determine whether $dE/dx$7 is experimentally accessible. In the CEPC drift-chamber design studied for machine-learning reconstruction, the detector has length about $dE/dx$8, radial extent from $dE/dx$9 to $\delta$0, about $\delta$1 layers, cell size $\delta$2, and a $\delta$3 He and $\delta$4 mixture. The corresponding waveform simulation uses Heed for ionization, Garfield++-based parameterization for transport and amplification, electronics response and noise from measurements, sampling at $\delta$5 over a $\delta$6 window, single-pulse rise time about $\delta$7, and noise about $\delta$8 (Tian et al., 2024).

Readout bandwidth, noise, and timing stability are recurrent limiting factors. The CERN H8 beam program for IDEA states that successful cluster counting required electronics with about $\delta$9 bandwidth and at least $dN/dx$0 sampling at $dN/dx$1 bits, implemented with a DRS-based system (Caputo et al., 2022). A later 120-channel CEPC-oriented drift-chamber prototype integrated a custom front-end with $dN/dx$2 waveform sampling and reported a $dN/dx$3 analog bandwidth of $dN/dx$4, an Equivalent Noise Input current of $dN/dx$5, and an intrinsic timing jitter of $dN/dx$6; cosmic-ray measurements showed that the system could resolve discrete ionization peaks within piled-up waveforms (Cai et al., 28 Mar 2026). An additional 24-channel ultra-low-noise preamplifier for drift-tube $dN/dx$7 measurements reached a bandwidth of $dN/dx$8, a charge gain of $dN/dx$9, an equivalent noise charge of $dE/dx$0, and a signal-to-noise ratio of $dE/dx$1 in He:iC$dE/dx$2H$dE/dx$3 (90:10), explicitly to preserve sub-fC, nanosecond-scale ionization structure (Ge et al., 20 May 2026).

High spatial granularity provides an alternative path to the same goal. In ILD TPC studies, cluster counting becomes feasible only when the readout pad size is sufficiently small that the spatial correlation of electrons from the same primary cluster survives drift and amplification; the simulation identifies a threshold below about $dE/dx$4, with especially noticeable improvement below about $dE/dx$5 (Einhaus et al., 2019). In the high-granularity CEPC TPC concept, $dE/dx$6 pads are small enough that each pad typically collects only a small number of electrons, allowing $dE/dx$7 reconstruction to be formulated as a point-cloud segmentation problem rather than a pure charge-summation task (Zhao et al., 12 Oct 2025).

3. Reconstruction algorithms: from peak finding to learned segmentation

Cluster counting is not identical to peak finding. In drift-chamber waveform analyses it is usually a two-stage reconstruction: first identify electron peaks in the waveform, then determine which peaks correspond to primary ionization clusters. Traditional methods are derivative-based. The DERIV family uses waveform amplitude together with first- and second-derivative thresholds; the RTA, or Running Template Algorithm, scans with an electron-pulse template and iteratively subtracts matched peaks; and higher-level CLUSTER procedures then merge consecutive-bin peaks and group temporally compatible electrons into primary clusters using diffusion-aware timing rules (D'Anzi et al., 2023).

A major practical complication is the mismatch between simulation and real data. A semi-supervised domain-adaptation approach treats peak finding as a binary classification problem on waveform segments of $dE/dx$8 bins, with simulation as the source domain and test-beam data as the target domain. The method uses optimal transport to align source and target samples and supplements the target domain with partial labels from a continuous wavelet transform. On pseudo-data, the reported AUC values are $dE/dx$9 for a source-trained baseline, $dE/dx$0 for unsupervised domain adaptation, $dE/dx$1 for semi-supervised domain adaptation, and $dE/dx$2 for the fully supervised ideal model; on CERN beam data the domain-adapted classifier shows better classification power than the traditional derivative-based algorithm and remains stable across varying track lengths (Zhao et al., 2024).

Machine-learning reconstruction in the CEPC drift chamber makes the two-stage structure explicit. The first stage is an LSTM peak finder operating on sliding windows of $dE/dx$3 points, with one LSTM layer of $dE/dx$4 hidden features, two fully connected layers, and a sigmoid output. At threshold $dE/dx$5, it reaches purity $dE/dx$6 and efficiency $dE/dx$7, while a second-derivative method matched to the same purity gives efficiency $dE/dx$8. The second stage is a DGCNN clusterizer in which detected peaks are graph nodes, peak times are node features, edges are built with $dE/dx$9-nearest neighbors in time, and $N$0 is the optimized setting. With an optimized threshold of $N$1, the method reconstructs cluster distributions close to MC truth and improves $N$2 separation by about $N$3 relative to the traditional algorithm (Tian et al., 2024).

For high-granularity TPCs the reconstruction target shifts from waveform peaks to hit-level primary-electron identification. The Graph Point Transformer (GraphPT) represents each track as a point cloud whose nodes carry charge and timing, connects nodes through Euclidean $N$4-nearest neighbors, and applies a graph-based U-Net with attention-based aggregation. The final $N$5 estimator thresholds per-hit probabilities and counts positive hits per unit track length. Relative to a truncated-mean baseline using activated pads per unit length, GraphPT improves hit-level classification from $N$6 accuracy/F1-score to $N$7 for the dot-product variant, and improves $N$8 separation by approximately $N$9 to $x$0 in the momentum interval from $x$1 to $x$2 (Zhao et al., 12 Oct 2025).

4. Simulation, calibration, and response modeling

Microscopic gas simulation underpins nearly all $x$3 studies. Garfield++ is commonly used as the reference model for primary ionization cluster formation, cluster-size distributions, electron drift, avalanche development, and induced signal formation. Because Geant4 does not provide the same microscopic cluster structure natively, parameterized interfaces have been developed to translate Geant4 energy-loss information into cluster-number and cluster-size distributions. One such program, built for IDEA-like drift chambers, studies a simplified chamber of $x$4 cells with $x$5 side, gas mixture $x$6 He + $x$7, and particle momenta from $x$8 to $x$9; it introduces three Geant4 algorithms for reproducing the cluster number and cluster size distributions, all consistent with Garfield++, with the third algorithm giving the closest match to the expected cluster-size shape (Cuna et al., 2021).

The IDEA simulation chain applies the same strategy at detector scale. Geant4 provides the deposited-energy history, Garfield++ supplies the cluster-statistics reference, and the conversion algorithm is constructed under the assumption of $\frac{dN}{dx}\approx \frac{N}{x}.$0 cluster-counting efficiency. In that idealized limit the fast Geant4-to-cluster model reproduces the cluster-number and cluster-size distributions and agrees well with fully microscopic Garfield++ simulations up to about $\frac{dN}{dx}\approx \frac{N}{x}.$1, although the Garfield++ result falls more rapidly at higher momentum (Caputo et al., 2022).

Not every $\frac{dN}{dx}\approx \frac{N}{x}.$2 analysis reconstructs clusters explicitly. In CEPC full-simulation PID studies, the TPC observable is modeled from Garfield++ lookup tables rather than extracted by a cluster-finding algorithm. The lookup tables provide the dN/dx mean and sigma as functions of particle angle and velocity, each gun configuration is fitted with a Gaussian, and track-level measurements are obtained by querying the tables and Gaussian-smearing the result. The observable is still interpreted explicitly as the number of initial ionization clusters per unit distance, but the analysis is response-based rather than waveform-based (Yu et al., 24 Jul 2025).

System-level physics studies may simplify the treatment further. In the $\frac{dN}{dx}\approx \frac{N}{x}.$3 forward–backward asymmetry analysis for future linear colliders, cluster counting is not reconstructed in the ILD chain; instead its expected effect is emulated by narrowing the TPC PID likelihood distributions. In the $\frac{dN}{dx}\approx \frac{N}{x}.$4 scenario the standard deviation of the $\frac{dN}{dx}\approx \frac{N}{x}.$5-distance is reduced by $\frac{dN}{dx}\approx \frac{N}{x}.$6, interpreted as about a $\frac{dN}{dx}\approx \frac{N}{x}.$7 improvement in separation, and a PerfectTPC benchmark reduces the width by $\frac{dN}{dx}\approx \frac{N}{x}.$8 (Márquez et al., 4 Mar 2026).

5. Empirical performance and detector-specific results

The reported gains from $\frac{dN}{dx}\approx \frac{N}{x}.$9 depend on gas mixture, detector topology, reconstruction algorithm, and the realism of the efficiency model, but the broad pattern is consistent: whenever primary-ionization structure is sufficiently preserved, cluster counting narrows the PID response relative to $dN_{\mathrm{cl}}/dx$00.

Detector or context	Implementation style	Reported outcome
IDEA drift chamber	Geant4-to-cluster model + H8 beam test	resolution 2 times better than traditional $dN_{\mathrm{cl}}/dx$01 (Caputo et al., 2022)
BESIII drift chamber	Garfield++/Heed + BOSS parameterization	at $dN_{\mathrm{cl}}/dx$02, $dN_{\mathrm{cl}}/dx$03 resolution about $dN_{\mathrm{cl}}/dx$04, $dN_{\mathrm{cl}}/dx$05 below $dN_{\mathrm{cl}}/dx$06 (Xin et al., 2022)
CEPC drift chamber	LSTM peak finder + DGCNN clusterizer	$dN_{\mathrm{cl}}/dx$07 separation improves by about $dN_{\mathrm{cl}}/dx$08; at $dN_{\mathrm{cl}}/dx$09, $dN_{\mathrm{cl}}/dx$10 vs $dN_{\mathrm{cl}}/dx$11 (Tian et al., 2024)
High-granularity CEPC TPC	GraphPT point-cloud segmentation	$dN_{\mathrm{cl}}/dx$12 separation improves by approximately $dN_{\mathrm{cl}}/dx$13 to $dN_{\mathrm{cl}}/dx$14 from $dN_{\mathrm{cl}}/dx$15 to $dN_{\mathrm{cl}}/dx$16 (Zhao et al., 12 Oct 2025)
ILD TPC	sub-$dN_{\mathrm{cl}}/dx$17 pad cluster counting	extrapolated $dN_{\mathrm{cl}}/dx$18 for $dN_{\mathrm{cl}}/dx$19 pads and $dN_{\mathrm{cl}}/dx$20 for $dN_{\mathrm{cl}}/dx$21 pads (Einhaus et al., 2019)
Helium-based beam studies	DERIV/RTA + waveform cleaning and correction	for $dN_{\mathrm{cl}}/dx$22 tracks, $dN_{\mathrm{cl}}/dx$23 resolution $dN_{\mathrm{cl}}/dx$24, $dN_{\mathrm{cl}}/dx$25 $dN_{\mathrm{cl}}/dx$26, improved to $dN_{\mathrm{cl}}/dx$27 after corrections (Elmetenawee et al., 26 Sep 2025)

For BESIII, the performance study is explicitly momentum-dependent. The $dN_{\mathrm{cl}}/dx$28 and $dN_{\mathrm{cl}}/dx$29 ionization curves cross around $dN_{\mathrm{cl}}/dx$30, so neither $dN_{\mathrm{cl}}/dx$31 nor $dN_{\mathrm{cl}}/dx$32 alone separates them well there, but above about $dN_{\mathrm{cl}}/dx$33 cluster counting shows a clear advantage. The ideal $dN_{\mathrm{cl}}/dx$34 case yields about a $dN_{\mathrm{cl}}/dx$35 improvement in $dN_{\mathrm{cl}}/dx$36 separation power over $dN_{\mathrm{cl}}/dx$37, and even with a $dN_{\mathrm{cl}}/dx$38 degradation in $dN_{\mathrm{cl}}/dx$39 resolution the gain remains about $dN_{\mathrm{cl}}/dx$40 (Xin et al., 2022).

The ILD TPC study establishes that cluster counting can match or exceed charge summation only in the high-granularity regime. With $dN_{\mathrm{cl}}/dx$41 tracks at $dN_{\mathrm{cl}}/dx$42, a separation power around $dN_{\mathrm{cl}}/dx$43 is obtained with $dN_{\mathrm{cl}}/dx$44 pads, and extrapolation to ILD-like track lengths of $dN_{\mathrm{cl}}/dx$45 gives $dN_{\mathrm{cl}}/dx$46 for $dN_{\mathrm{cl}}/dx$47 pads and $dN_{\mathrm{cl}}/dx$48 for $dN_{\mathrm{cl}}/dx$49 pads. The same study finds that cluster counting surpasses the maximum separation power of charge summation for drift lengths above about $dN_{\mathrm{cl}}/dx$50 (Einhaus et al., 2019).

Early full-length drift-chamber prototype data already showed the complementarity of charge and cluster observables. In the TRIUMF measurements with a $dN_{\mathrm{cl}}/dx$51 beam, cluster counting combined with a $dN_{\mathrm{cl}}/dx$52 truncated-mean charge measurement improved pion selection efficiency at $dN_{\mathrm{cl}}/dx$53 muon rejection by about $dN_{\mathrm{cl}}/dx$54 percentage points, with a typical example from about $dN_{\mathrm{cl}}/dx$55 to $dN_{\mathrm{cl}}/dx$56. That study also found optimal results for a signal smoothing time of $dN_{\mathrm{cl}}/dx$57, corresponding to a $dN_{\mathrm{cl}}/dx$58 Nyquist frequency (Caron et al., 2013).

6. Inefficiencies, detector effects, and hybrid PID strategies

The central limitation of $dN_{\mathrm{cl}}/dx$59 is that fully efficient cluster counting is not achievable in practice. Multiple studies identify space charge, electron attachment, and recombination as the dominant loss mechanisms. In the IDEA beam program these effects reduce the effective cluster-counting efficiency to about $dN_{\mathrm{cl}}/dx$60 in He/iC$dN_{\mathrm{cl}}/dx$61H$dN_{\mathrm{cl}}/dx$62 90/10 (Caputo et al., 2022). In real waveform analyses, the raw number of detected clusters also decreases with drift time; one beam-test study reports a cluster loss of about $dN_{\mathrm{cl}}/dx$63 clusters every $dN_{\mathrm{cl}}/dx$64, attributes it to recombination, attachment, and electric-field suppression near the sense wire, and notes that the correction is geometry-dependent rather than universal (D'Anzi et al., 2023).

Temporal overlap is the second major limitation. The method becomes less favorable when clusters overlap too strongly or when the ionization pattern is too dense. In the 2025 CERN beam studies, normal incidence in the 80/20 He–isobutane mixture produces a counting deficit attributed to higher local cluster density, space-charge effects, and increased overlap in time. The same study also shows that gas gain, impact parameter, and track angle all affect counting efficiency, and applies waveform cleaning together with a recombination/attachment correction

$dN_{\mathrm{cl}}/dx$65

to restore the expected Poisson-like behavior (Elmetenawee et al., 26 Sep 2025).

These limitations make hybrid PID particularly important. IDEA simulation finds particularly good $dN_{\mathrm{cl}}/dx$66 separation over the full momentum range except roughly $dN_{\mathrm{cl}}/dx$67, where an additional time-of-flight measurement with about $dN_{\mathrm{cl}}/dx$68 resolution over a $dN_{\mathrm{cl}}/dx$69 path would recover the separation (Caputo et al., 2022). BESIII studies report an analogous crossover region around $dN_{\mathrm{cl}}/dx$70–$dN_{\mathrm{cl}}/dx$71, where TOF raises efficiency from about $dN_{\mathrm{cl}}/dx$72 to $dN_{\mathrm{cl}}/dx$73 near $dN_{\mathrm{cl}}/dx$74 (Xin et al., 2022).

At CEPC, full-event studies make the case for explicit $dN_{\mathrm{cl}}/dx$75+ToF combination. A TPC-only dN/dx strategy is highly efficient for kaons, with efficiency $dN_{\mathrm{cl}}/dx$76, but purity is only $dN_{\mathrm{cl}}/dx$77 because of severe pion contamination. Adding OTK time of flight raises purity to $dN_{\mathrm{cl}}/dx$78; combining ITK, TPC, and OTK yields $dN_{\mathrm{cl}}/dx$79 efficiency and $dN_{\mathrm{cl}}/dx$80 purity; and a momentum-dependent hybrid strategy reaches $dN_{\mathrm{cl}}/dx$81 efficiency and $dN_{\mathrm{cl}}/dx$82 purity (Yu et al., 24 Jul 2025). This suggests that, in realistic hadronic environments, $dN_{\mathrm{cl}}/dx$83 is best understood as a high-precision ionization observable whose full utility often depends on complementary timing information.

The same conclusion appears at analysis level. In the $dN_{\mathrm{cl}}/dx$84 forward–backward asymmetry study, the emulated $dN_{\mathrm{cl}}/dx$85 scenario sharpens kaon identification by narrowing the effective PID width by $dN_{\mathrm{cl}}/dx$86, corresponding to about a $dN_{\mathrm{cl}}/dx$87 improvement in $dN_{\mathrm{cl}}/dx$88 separation, and thereby reduces the uncertainty on $dN_{\mathrm{cl}}/dx$89 relative to standard $dN_{\mathrm{cl}}/dx$90 (Márquez et al., 4 Mar 2026).

7. Scalable implementations and emerging directions

Recent work has increasingly treated cluster counting not only as a reconstruction problem but also as a readout-architecture problem. Edge machine-learning studies for next-generation drift chambers replace explicit peak finding and clusterization with direct regression from waveform to cluster count. In one CEPC-inspired setup, the input is a waveform truncated to $dN_{\mathrm{cl}}/dx$91 samples, the baseline model is a fully connected DNN with architecture $dN_{\mathrm{cl}}/dx$92, the labels are Garfield++ truth cluster counts, and the projected $dN_{\mathrm{cl}}/dx$93-track performance exceeds $dN_{\mathrm{cl}}/dx$94 pion/kaon separation across $dN_{\mathrm{cl}}/dx$95 to $dN_{\mathrm{cl}}/dx$96. When synthesized with hls4ml, the baseline model reaches $dN_{\mathrm{cl}}/dx$97 latency, a quantized $dN_{\mathrm{cl}}/dx$98 version also reaches $dN_{\mathrm{cl}}/dx$99, and a $dE/dx$00-pruned quantized model reaches $dE/dx$01 (Yilmaz et al., 13 Nov 2025).

Dedicated electronics programs indicate that such architectures are no longer purely conceptual. The modular $dE/dx$02-channel readout prototype for CEPC $dE/dx$03 studies and the 24-channel ultra-low-noise preamplifier for drift-tube detectors both demonstrate that bandwidth, noise, timing, and channel density can be pushed into the regime required for resolved-cluster measurements (Cai et al., 28 Mar 2026, Ge et al., 20 May 2026).

Bandwidth requirements, however, are not universal. A full-length prototype study from 2013 concluded that cluster counting did not require an overly high sampling rate and found optimal results with $dE/dx$04 smoothing, corresponding to a $dE/dx$05 Nyquist frequency (Caron et al., 2013). Later helium-based drift-chamber programs instead emphasized about $dE/dx$06 bandwidth and at least $dE/dx$07 sampling for successful cluster counting (Caputo et al., 2022). A plausible implication is that the front-end requirement is architecture-dependent: combined charge-plus-cluster discriminants, explicit single-electron peak reconstruction, and high-granularity hit-level segmentation do not demand identical signal fidelity.

Taken together, these developments place $dE/dx$08 at the intersection of gas microphysics, waveform inference, detector segmentation, and front-end electronics. The technique is no longer confined to idealized counting arguments; it now includes Geant4-to-Garfield++ response transfer, domain adaptation for simulation–data mismatch, graph-based reconstruction in TPCs, and FPGA-compatible inference for real-time readout. The remaining open problems are correspondingly practical: calibration of detector-specific losses, validation beyond simulation, control of overlap and gain dependence, and integration of $dE/dx$09 with timing and tracking in full collider environments.