Particle-Flow Reconstruction in Collider Physics

• Particle-flow reconstruction is a method for integrating diverse detector signals to accurately identify and reconstruct stable particles.
• It combines tracking, calorimetry, and muon detection to improve jet energy and missing transverse momentum resolutions by up to 3x over calorimeter-only approaches.
• Modern implementations leverage machine learning architectures like graph neural networks and transformers to boost efficiency, resolution, and scalability in high-energy physics experiments.

Particle-flow reconstruction is a paradigm for identifying, classifying, and reconstructing the complete set of stable particles produced in high-energy physics detectors by combining information from all available subdetector systems on an event-by-event basis. The dominant context remains collider environments, particularly hadron colliders such as the LHC, where particle flow (PF) algorithms have revolutionized object resolutions for jets, missing transverse momentum ($p_T^{\mathrm{miss}}$), electrons, muons, and hadronic taus. In recent years, PF methodology has expanded into liquid- and gas-phase flow diagnostics, highly granular imaging calorimetry, and, notably, unified detector simulation and reconstruction pipelines utilizing advanced machine-learning architectures (Collaboration, 2017, Beaudette, 2014, Collaboration, 24 Jan 2026, Mokhtar, 28 Aug 2025, Pata et al., 2023, Kakati et al., 2024, Dreyer et al., 25 Mar 2025, Dreyer et al., 2024, Nguyen, 2011, Bello et al., 2020, Zhou et al., 2023, Lasinger et al., 2018).

1. Conceptual Basis and Physical Motivations

Particle-flow reconstruction solves a fundamental inference problem: from indirect, overlapping, and incomplete subdetector measurements—charged particle trajectories (tracker), electromagnetic/hadronic calorimeter energy depositions, and muon hits—produce an optimal, self-consistent list of final-state particles with four-vectors and species assignments. This approach leverages the optimal subsystem for each observable: track $p_T$ resolution (CMS: $\sigma_{p_T}/p_T \sim 1–2\%$ for $p_T \lesssim 100\,\mathrm{GeV}$), ECAL granularity and energy precision for photons and electrons ($\sim2\%$ for barrel ECAL), and the wide HCAL acceptance for neutral hadrons (Collaboration, 2017).

Central motivations include:

• Factor $\sim1.5–3$ improvements in jet and missing $p_T$ resolutions versus calorimeter-only methods;
• Robustness to jet fragmentation spectrum, reducing quark/gluon response differences by $\sim2\times$ compared to calorimetric jets;
• Particle-level pileup mitigation (e.g., charged hadron subtraction) and lepton isolation stable under increasing pileup (Collaboration, 2017, Beaudette, 2014, Nguyen, 2011).

In heavy-ion and high-pileup environments, PF reconstruction can uniquely exploit tracking to suppress non-linear HCAL systematics and provides fine angular resolution ($\sim$0.01\,rad at 100\,GeV), critical for jet quenching and substructure studies (Nguyen, 2011).

2. Canonical Algorithms: Object Linking and Classification

Historical PF algorithms in collider contexts proceed in a modular fashion:

1. Low-level object reconstruction: Iterative Kalman-filter track finding; calorimeter and preshower topological clustering; muon reconstruction in standalone, global, and tracker seeding modes (Collaboration, 2017).
2. Link-building ("block" construction): Deterministic linking of tracks to calorimeter clusters via track extrapolation, geometric/spatial overlap, and energy compatibility. ECAL and HCAL clusters are cross-linked, and clusters associated to tracks are grouped in shared blocks (Collaboration, 2017, Beaudette, 2014).
3. Block hypothesis and particle identification: Ordered logic assigns PF muons, electrons (with GSF tracks and ECAL superclusters), photons (unlinked ECAL clusters), charged hadrons (track-linked clusters), and neutral hadrons (unlinked HCAL clusters). Particle energies are assigned based on a hierarchy of calibrations, e.g., tracker $p$ for charged hadrons, energy-weighted sums for neutral hadrons (Beaudette, 2014).
4. Consistency and ambiguity resolution: Direct comparison between sum of tracks and calorimeter energy after per-particle calibration, with χ²-based (or expectation-maximization for overlapping clusters) hypothesis test. Substantial residuals trigger candidate energy sharing, splitting, or further ambiguity-resolving procedures (Collaboration, 2017).

Variant implementations such as PandoraPFA introduce per-hit local software compensation, adjusting hit contributions according to local energy density via a phenomenologically motivated weight function, calibrated using single hadron samples. This is critical for correcting non-compensating calorimeter response and further improves jet energy resolution by $10–20\%$ (Tran et al., 2017).

Table 1. Core stages of canonical PF algorithms

Stage	Methodology	Function
Local object building	Kalman-filter tracking, topocluster	Localizes inputs
Linking	Geometric/energy sharing & EM	Groups related signals
Particle ID	Logical/test, boosted decision trees	Species and kinematic assignment
Final assembly	χ² and expectation-maximization fits	Mitigates ambiguities

For heavy-ion (PbPb) collisions, CMS adapts by enforcing tighter pixel-only track seeding, lowering tracking efficiency to $60–70\%$ in 0–10% central events, but maintaining fake rates < few percent. Pre-subtraction of diffuse underlying event from calorimeter towers is also standard (Nguyen, 2011).

3. Machine-Learned PF: Set-to-Set Inference, Architectures, and Losses

Modern PF research recasts the problem as a supervised set-to-set mapping from tracks/clusters to reconstructed particles, optimized by learning architectures. Numerous successful instantiations exist:

• Graph Neural Networks (GNNs): MLPF (CMS and Fermilab) uses message-passing on kNN graphs built dynamically in learnable embedding spaces. Each node (input track/cluster) is paired with a classification head (particle type, existence) and a regression head (five-vector). MLPF achieves competitive or improved jet $p_T$ resolution (up to 50% reduction in interquartile range) vs rule-based PF (Pata et al., 2021, Pata et al., 2023, Pata et al., 2022).
• Transformers: Set-to-set architectures avoid explicit graph construction by leveraging multi-head self-attention, reaching $\mathcal{O}(N)$ scaling using kernel approximations (linear transformers). CMS MLPF (2026) employs dual encoder transformer blocks with separate heads for primary object classification, particle ID, and four-momentum regression, yielding up to 20% improvement in jet energy resolution at 30–100\,GeV (Collaboration, 24 Jan 2026, Mokhtar, 28 Aug 2025).
• Hypergraph Assignment Models: HGPflow formulates PF as a hypergraph assignment, where particle candidates correspond to hyperedges linking sets of input objects, trained with a combination of assignment loss (object condensation style) and kinematic regression. This scheme surpasses both MLPF and rule-based PF in efficiency, purity, and jet energy resolution (Kakati et al., 2024).

The canonical per-element loss in such models is

$$L = \sum_{i} \Bigl[ L_\mathrm{cls}(\text{type}_i, \text{type}_i') + \alpha L_\mathrm{reg}(\vec{p}_i, \vec{p}_i') \Bigr]$$

where $L_\mathrm{cls}$ is typically a (focal) cross-entropy and $L_\mathrm{reg}$ a $p_T$-weighted MSE or Huber loss, with careful weighting to match target physics resolutions (Mokhtar, 28 Aug 2025, Pata et al., 2023).

Transfer learning methods (fine-tuning large backbone transformer models pretrained on one detector geometry for another) accelerate convergence and reduce required samples by an order of magnitude, matching rule-based PF performance after 100k events rather than 1M (Mokhtar et al., 28 Feb 2025).

4. Specialized and Extended PF: Imaging, Generative Surrogates, Fluid Flow

Beyond standard tracking–calorimeter combinatorics, several workstreams extend PF to non-standard domains:

• Computer Vision PF: Calorimeter data is directly represented as multi-channel images; deep convolutional/graph networks (U-Net, Dynamic Graph CNN, DeepSet) are trained to assign cell-level energy fractions between neutral and charged contributions. This image-based regression achieves up to 4× better neutral energy resolution than parametric PF in regions of high signal overlap, and super-resolution networks can upsample coarse input images (e.g. 8×8 to 32×32) to recover high-fidelity neutral shower topology (Bello et al., 2020).
• Score-Based Generative Models ("PF Surrogates"): Parnassus and its full-event extension replace explicit detector simulation + PF chain. A conditional normalizing flow or diffusion model, conditioned on truth-level particle sets, maps to reconstructed PF objects, including kinematics, particle type, and vertex. This pipeline replicates both particle- and jet-level resolutions of CMS PF (p_T, η, φ, multiplicity, substructure) with negligible bias across a wide range of physics processes, and natively models mis-ID, pileup fakes, and event topology (Dreyer et al., 2024, Dreyer et al., 25 Mar 2025).
• Lagrangian Particle-Flow in Fluids: In experimental fluid mechanics, "particle-flow reconstruction" refers to approaches that integrate sparse Lagrangian trajectories and dense Eulerian flow fields within a unified physics-informed optimization (jointly fitting neural PDE surrogates and data-constrained polynomials or point sets), yielding velocity and pressure fields consistent with governing equations as well as fitted particle properties (Zhou et al., 2023, Lasinger et al., 2018).

5. Performance Benchmarks, Scalability, and Physics Impact

Consistent findings across collider PF variants:

• Jet energy resolution is improved by $1.2–1.5\times$ versus calorimeter jets across $p_T=30–1000\,\mathrm{GeV}$; up to 50% (IQR) reduction for the best GNN/transformer models (Collaboration, 24 Jan 2026, Pata et al., 2023, Mokhtar et al., 28 Feb 2025, Nguyen, 2011).
• MET and lepton isolation are more robust, especially in high-pileup.
• Pileup subtraction and mitigation can operate at the particle (candidate) level, preserving reconstruction performance at $\mu\gtrsim 20–100$.
• Inference times for end-to-end ML models are consistently under 50 ms per full event (O(10k) elements) on commodity GPUs, scaling linearly with event size (Collaboration, 24 Jan 2026, Pata et al., 2022, Pata et al., 2023).

MLPF performance in CMS and e⁺e⁻ detectors at both particle and event level is quantifiably equal or superior to rule-based PF—e.g., neutral hadron efficiency is improved ($\epsilon>90\%$) with reduced fake rates, jet response is unbiased and sharper, and generalization to new detector geometries or physics processes is robust (Pata et al., 2021, Mokhtar et al., 28 Feb 2025, Collaboration, 24 Jan 2026).

Full-event generative models achieve sub-percent agreement on all core physics observables with full simulation+PF and significantly outperform Delphes across kinematic, multiplicity, and substructure variables (Dreyer et al., 25 Mar 2025, Dreyer et al., 2024).

6. Open Frontiers and Future Directions

Active research aims to:

• Elevate PF from hand-crafted object-centric algorithms to fully end-to-end differentiable, trainable models, integrating both upstream low-level (tracking, clustering) and downstream (jet, MET, tagger) tasks (Collaboration, 24 Jan 2026, Mokhtar, 28 Aug 2025).
• Incorporate timing information, point-cloud and hypergraph models, and ultra-high granularity calorimetry (HGCAL).
• Develop "foundation models" for PF suitable for cross-collider and cross-detector applications, leveraging transfer learning and self-supervised techniques (Mokhtar et al., 28 Feb 2025).
• Enable real-time, trigger-level, or FPGA/edge deployment via quantization, pruning, and hardware-optimized architectures (Pata et al., 2023, Pata et al., 2022).
• Unify simulation and reconstruction into single generative surrogates for fast analysis turn-around and phenomenology (Dreyer et al., 25 Mar 2025, Dreyer et al., 2024).
• Extend to highly dense environments (extreme pileup, high-energy heavy-ion) via ever-more robust assignment and uncertainty-aware or probabilistic procedures [(Elagin et al., 2012) (abstract only)].

The continued evolution of particle-flow reconstruction is now converging on highly scalable, data-driven, and physically-informed deep-learning architectures, with demonstrated impact on both the precision and efficiency of global event interpretation as well as the practical computational throughput required for modern and future collider physics.