Pile-Up Simulation and Mitigation

• Pile-Up Simulation and Mitigation is the process of modeling overlapping events within a detector and applying corrective methods to recover true signal properties.
• Techniques involve Poisson process modeling, Fourier inversion, and Monte Carlo simulations to address energy bias, rate suppression, and merged event issues.
• Practical solutions such as area-median subtraction, per-particle corrections, and machine learning enhance data accuracy in high-rate experimental environments.

Pile-up refers to the superposition of multiple independent events within the spatial or temporal resolution element of a detector, such that the individual events cannot be distinguished as separate by the measurement system. This phenomenon, ubiquitous in high-luminosity colliders, X-ray CCDs, photon-counting detectors, microdosimetric measurements, and bolometric or pulse-height applications, results in severe distortions of reconstructed observables: total event rates are suppressed, energy and momentum spectra are biased, and physically meaningful signatures may be obscured or rendered intractable. Modern simulation frameworks and data analysis pipelines must model, simulate, and mitigate pile-up effects both to extract unbiased physics from current experiments and to guarantee performance at future ultra-high-rate facilities.

1. Physical Origins and Statistical Description of Pile-Up

The fundamental characteristic of pile-up is the registration of two or more physically distinct events (particles, photons, energy depositions) as a single detected event due to the finite time or spatial granularity of the instrument. In collider detectors, pile-up arises from multiple proton-proton interactions per bunch crossing (typical ⟨μ⟩=20–200 at the LHC/HL-LHC). In X-ray astronomy, pile-up manifests in CCDs when two or more photons interact within the same pixel or adjacent pixels inside a readout frame, leading to misclassification or loss (“grade migration”) (Yoneyama et al., 25 Nov 2025). In single-photon counting modules, electronic dead time enforces a minimal resolvable interval, with arrivals within this interval subject to merging or loss (&&&1&&&). Microdosimetric devices such as TEPCs register multiple traversals as a single high-deposition event if their signal shaping times are exceeded (Pierobon et al., 4 Apr 2025).

Statistically, pile-up is governed by Poisson processes. Denote λ as the mean rate of incident events per resolution element (time/window/pixel) Δ. The per-resolution-cell statistics follow the Poisson law:

$$P(k; \lambda) = \frac{\lambda^k e^{-\lambda}}{k!} \,,\ \ \ (k=0,1,2,\ldots)$$

If $k \geq 2$ events occur within Δ, their attributes (energies, momenta) are summed and registered as a single event. As a consequence, the observed (apparent) rate distribution undercounts the true rate in the single-event region and exhibits a distortion at summed attribute values. The first moment (rate × attribute) is conserved, but total counts are always underestimated (Deshpande et al., 2012).

2. Simulation of Pile-Up in Particle Physics and Astrophysics

Collider Experiments

In hadron collider simulations, event generators (e.g., Pythia, Sherpa, Powheg) produce both hard-scatter and minimum-bias (pile-up) event samples. For each event, a Poisson-distributed number of pile-up interactions (with mean μ) are overlaid with the primary hard interaction, and all subsequent detector effects—tracking, calorimetry, energy smearing—are simulated (often via GEANT4 or Delphes) (Collaboration, 2015, Collaboration, 2017). Such simulation chains allow study of object performance (jet pT, MET, substructure, isolation) at pile-up levels up to μ ≈ 140–200 (Maier et al., 2021).

X-ray and Photon-Counting Detectors

CCDs and photon-counting sensors are modeled by assigning each primary photon a random arrival time and impact position. The merging logic—merging photons within the same pixel or event-recognition cell per frame—is implemented either via analytic Poisson calculations (Yoneyama et al., 25 Nov 2025), Monte Carlo tracking (Geant4-based; ComptonSoft for XRISM) (Yoneyama et al., 25 Nov 2025), or full waveform simulation (Ahsan et al., 2022). In the case of SPACIROC-3 MAPMTs, electronic dead time is applied by discarding or merging close-arrival pulses, and the resulting counts-per-frame statistics are captured via renewal theory and explicit Poisson–dead-time models (M'sihid et al., 28 Nov 2025).

Microdosimetry and Bolometers

Pile-up in microdosimeters and bolometers is simulated by taking a sequence of single-event depositions and overlaying events probabilistically according to the measured or known event rate and the instrument's temporal resolution or shaping time. The final spectrum is then a mixture of single and convolved (pile-up) events (Pierobon et al., 4 Apr 2025, Armatol et al., 2020).

3. Analytical and Algorithmic Correction of Pile-Up Effects

A comprehensive set of analytical and data-driven techniques have emerged for both determining and mitigating pile-up distortions:

Generalized Analytical Inversion

If the pile-up can be described as Poisson merging of events, the true differential rate distribution λ(S) can be recovered from the observed (pile-up distorted) apparent distribution λ_a(S) by Fourier transform methods:

$$\tilde{P}^a(f) = \exp\left( -\Lambda_0 \right) \exp[\tilde{P}(f)],\quad \ln \tilde{P}^a(f) = \tilde{P}(f) - \Lambda_0$$

where $\tilde{P}^a(f)$ is the Fourier transform of the apparent probability density, $\tilde{P}(f)$ that of the true distribution, and $\Lambda_0 = \sum_i \lambda(S_i)$ (Deshpande et al., 2012). The inversion is performed via forward and inverse FFTs, with log-domain Hadamard corrections and regularization for stability.

Pile-Up–Aware Statistical Models in Binned Data

Modified Poisson models incorporate the conditional saturation or merging probability per count/bin, introducing an extra pile-up parameter ρ, yielding closed forms for the count and waiting-time distributions suitable for likelihood inference or batch deconvolution (Sevilla, 2013).

Monte Carlo–Based Library and Deconvolution Approaches

In SDD, CCD, or microdosimeter contexts, simulation-based approaches precompute large Monte Carlo libraries of detector responses for all possible input energies and merging scenarios (including merged pulse shapes and "grade migration"). During analysis, the observed distribution is then matched via iterative fitting and forward-model sampling, optionally using maximum-likelihood or Cash-statistic optimization (Tamba et al., 2021, König et al., 14 Nov 2025, Ahsan et al., 2022, Pierobon et al., 4 Apr 2025).

4. Pile-Up Mitigation Techniques in Collider Physics

Mitigating pile-up in high-luminosity colliders is accomplished through both data-driven and AI-based approaches:

Area–Median Subtraction

The jet area/median method corrects the pile-up-induced shift in jet (or object) momentum by subtracting the event-by-event median energy density ρ times the jet area A_j:

$$p_T^{\text{sub},j} = p_T^{\text{raw},j} - A_j \,\rho$$

with ρ estimated as the median $p_T/A$ over soft jets, and A_j obtained from ghost-inclusion techniques in FastJet (0707.1378, Soyez, 2018). The method is robust, data-driven, and general across jet algorithms.

Particle-Level Per-Object and Per-Particle Correction

The CMS Charged Hadron Subtraction (CHS) algorithm removes all charged PF candidates consistent with pile-up vertices and performs area subtraction for neutral pile-up (Collaboration, 2020). Pileup Per Particle Identification (PUPPI) generalizes by calculating a per-particle local activity metric α_i, referencing event-wide pile-up distributions (from charged pile-up proxy), and assigning a continuous weight:

$w_i = F_{\chi^2_1}\left(\chi_i^2\right)$

where

$$\chi_i^2 = \Theta(\alpha_i - \bar{\alpha}_{PU})\frac{(\alpha_i - \bar{\alpha}_{PU})^2}{\sigma^2_{PU}}$$

and the four-momentum is rescaled by $w_i$ (Bertolini et al., 2014, Collaboration, 2020). Constituent-level subtraction, grooming (trimming, soft drop, pruning), and jet-cleansing are also employed to remove residual pile-up (Collaboration, 2015, Soyez, 2018).

Machine-Learned Pile-Up Suppression

Recent developments leverage AI architectures:

• Transformer-based regressors (PUMA, PUMiNet) operate at the particle or jet level using event-wide, banded or global attention to regress energy- or mass-fraction or energy-fraction weights for each reconstructed object. PUMA uses sparse transformers with hierarchical clustering and attention, yielding improved MET and dijet mass resolutions at μ ≈ 140–200, surpassing classical algorithms (Maier et al., 2021, Vaughan et al., 4 Mar 2025).
• Graph Neural Networks (GGNN) trained on extended topological and local-density features (using PUPPI-style α_i, neighbor connectivity, and global densities) outperform rule-based methods in pile-up rejection and reconstructed jet resolution (Martinez et al., 2018).

Forward Region Strategies

In regions lacking tracking, ATLAS has pioneered the use of jet timing, width, and core energy–fraction (γ-fit) alongside topological forward Jet Vertex Tagging (fJVT) that exploits per-vertex missing transverse momentum signatures to reject pile-up jets, yielding 49–67% background rejection at 85% hard-scatter efficiency in the forward region for typical μ ≈ 22–35 (Collaboration, 2017).

Wavelet-Based Filtering

Pile-up is treated as spatial "white noise" in the rapidity–azimuth ($y,\phi$) plane, and events are filtered by band-dependent thresholding in the wavelet domain (wavelet coefficients $C_{j,k}$ hard-thresholded according to the noise scaling law $\sigma_j \sim \sqrt{\mu}$). The method can be implemented in FPGA or GPU with L1 trigger feasibility (Monk et al., 2018).

5. Pile-Up in X-ray CCDs and Spectroscopic Correction

Pile-up in photon-counting devices leads to undercounting, spectral hardening, and grade migration (multi-photon events fall outside valid event patterns):

• The pile-up fraction for a cell with mean μ per frame is $P_\text{pile} = 1 - e^{-\mu} (1+\mu)$ (Yoneyama et al., 25 Nov 2025).
• Nonlinear spectral responses are accurately modeled by Monte Carlo charge transport, frame readout, and event grading (Tamba et al., 2021) and corrected via sampling-based spectral fitting with detector response linbraries.
• For moderate count rates, first-principles two-photon convolution models yield analytic correction formulas for the spectrum, using explicit integration over the pulse formation time and account for the probability of one- versus two-photon pile-ups in the measurement window (Ahsan et al., 2022).
• At high rates, normalizing flows trained on stacks of simulated annular spectra enable posterior inference of physical source parameters (e.g., flux, temperature, absorption), with uncertainty calibration matching or exceeding that of traditional excision-based MCMC (König et al., 14 Nov 2025).

6. Pile-Up Estimation and Correction in Microdosimetry, Bolometric, and Photon-Counting Modes

Microdosimetry

Pile-up introduces a convolution distortion of the microdosimetric lineal-energy distribution. A stochastic matching algorithm (using the Kolmogorov–Smirnov test versus GEANT4-based reference libraries) determines the effective pile-up probability $p(R)$ as a function of experimental rate $R$, typically linear for $R \lesssim 3 \times 10^4$ pps, with a saturation at higher rates (Pierobon et al., 4 Apr 2025). Corrective deconvolution, or first-order inversion in the small-pileup limit, restores observables to robust accuracy (<5% bias) for $p<0.05$.

Bolometers and Pulse-Counting

Controlled dual-pulse tests with programmable waveforms allow calibration of discrimination power and rejection efficiency for heating events separated by intervals as low as 2 ms (Armatol et al., 2020). Optimum filtering, rise/delay–shape analysis, and event-by-event statistical cuts achieve >90% pile-up rejection down to $\Delta t_{90}\sim 2$ ms.

Electronics Dead-Time Pile-Up

In fast counting systems with known double-pulse resolution $\tau$, the relation

$$R_\mathrm{obs} = R_\mathrm{true} \, e^{-R_\mathrm{true}\tau}$$

connects measured and true rates and can be inverted analytically using the Lambert–$W$ function. Per-channel dead time is estimated via machine learning on per-pixel count histograms, and uncertainty propagation quantifies statistical and systematic errors (M'sihid et al., 28 Nov 2025).

7. Practical Implementation, Validation, and Limitations

Implementation of pile-up mitigation in modern experiments requires:

• Precise and stable estimation of key instrument parameters: e.g., jet area and event density for area subtraction, per-pixel dead times, event-shape thresholds.
• Regularization and windowing for numerical stability in analytic inversion (e.g., Wiener filter in the Fourier-log domain) (Deshpande et al., 2012).
• Large-scale Monte Carlo libraries for fast forward-model spectral analysis (Tamba et al., 2021, König et al., 14 Nov 2025).
• Machine learning architectures deployed on CPU/GPU hardware with O(10 ms/event) latency for transformer-based solutions at $N_\mathrm{particles}\sim 9\times 10^3$ per event (Maier et al., 2021).
• Careful trade-offs among bias, resolution stability, and computational efficiency in the choice of method (e.g., PUPPI vs. area-median vs. wavelet).
• Full data/simulation closure checks, empirical cross-validation with control channels (e.g., Z+jets for jet performance), and propagation of uncertainties from pile-up modeling to physics observables (Collaboration, 2015, Collaboration, 2017, Tamba et al., 2021, Yoneyama et al., 25 Nov 2025, Pierobon et al., 4 Apr 2025).

Known limitations include the assumption of uncorrelated Poissonian pile-up, possible residual biases in highly inhomogeneous backgrounds, statistical uncertainty at large bins/S regions, saturation or ill-conditioning at extreme rates, and the need for regular re-calibration. Extensions incorporating time/veto information, dynamic event-wide context, and real-time edge computation are active research areas (Vaughan et al., 4 Mar 2025, M'sihid et al., 28 Nov 2025).

---

Table: Key Pile-Up Mitigation Techniques and Representative Applications

Technique	Underlying Principle	Representative References
Area-median subtraction (ρA)	Jet-wise pT density & area correction	(0707.1378, Soyez, 2018, Collaboration, 2015)
PUPPI (per-particle ID)	Local activity + pile-up proxy weights	(Bertolini et al., 2014, Collaboration, 2020)
Machine learning (Transformer/GNN)	Event/contextual per-object regression	(Maier et al., 2021, Vaughan et al., 4 Mar 2025, Martinez et al., 2018)
Forward tagging and fJVT	Timing, shape, topology discrimination	(Collaboration, 2017)
Wavelet decomposition	Multiscale noise-domain filtering	(Monk et al., 2018)
Library-based spectral correction	MC sample-based non-linear analysis	(Tamba et al., 2021, König et al., 14 Nov 2025)
Statistical/analytical inversion	Poisson/log-Fourier, deconvolution	(Deshpande et al., 2012, Sevilla, 2013, Ahsan et al., 2022)
Dead-time inversion + ML	Analytical inversion, ML dead-time est	(M'sihid et al., 28 Nov 2025)

The field continues to evolve rapidly, with future pile-up mitigation strategies integrating deep learning, hardware-embedded real-time processing, full event-wide combinatorics, and adaptive, data-driven calibration for next-generation high-rate detectors across diverse domains.