Mlpf Particle Flow Eval

---
name: mlpf-particle-flow-eval
description: Evaluates a graph neural network's ability to reconstruct particle-flow objects (charged and neutral hadrons) from detector-level tracks and calorimeter clusters in high-pileup simulated events. It probes multi-task learning for particle classification and momentum/energy regression under realistic collider conditions. Use when the user wants to benchmark on DELPHES simulated QCD multijet and ttbar events, or asks about evaluating this task. Reports Efficiency.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2101.08578
  bibtex_key: pata2021mlpf
  confidence: high
---

# mlpf-particle-flow-eval

> MLPF: Efficient machine-learned particle-flow reconstruction using graph neural networks — Pata et al. (2021) (arXiv:2101.08578, 2021)

## What this evaluates

Evaluates a graph neural network's ability to reconstruct particle-flow objects (charged and neutral hadrons) from detector-level tracks and calorimeter clusters in high-pileup simulated events. It probes multi-task learning for particle classification and momentum/energy regression under realistic collider conditions.

## Datasets

- **DELPHES simulated QCD multijet and ttbar events** — total ?; splits: train (-1), val (-1); repo https://github.com/jpata/particleflow

## Metrics

- `Efficiency` **(primary)** — range: [0, 1]
  - True positive rate for particle identification, calculated as TP / (TP + FN). Measures the fraction of true particles correctly reconstructed.
- `Fake rate` — range: [0, 1]
  - False positive rate for spurious particle reconstruction, calculated as FP / (FP + TN). Measures the fraction of reconstructed particles that are not present in the truth.
- `pT(E) resolution` — range: other
  - Standard deviation of the relative difference between predicted and true transverse momentum or energy, typically std((p_pred - p_true) / p_true).
- `η resolution` — range: other
  - Standard deviation of the pseudorapidity prediction error, std(η_pred - η_true).
- `N resolution` — range: other
  - Standard deviation of the predicted versus true particle multiplicity per event, std(N_pred - N_true).

## Input / output format

**Input**: Graph-structured detector data containing tracks and calorimeter clusters for each event, representing a high-pileup environment.

**Output**: Per-particle multi-class labels (e.g., charged hadron, neutral hadron) and continuous regression targets for transverse momentum (pT), energy (E), and pseudorapidity (η).

## Scoring recipe

```python
def compute_metrics(predictions, truth):
    # Classification metrics per particle class
    tp = sum(pred == true & true == class_label)
    fn = sum(pred != true & true == class_label)
    fp = sum(pred == class_label & true != class_label)
    tn = total_negatives - fp
    efficiency = tp / (tp + fn)
    fake_rate = fp / (fp + tn)
    
    # Regression resolutions
    pt_res = std((p_pred - p_true) / p_true)
    eta_res = std(eta_pred - eta_true)
    n_res = std(N_pred - N_true)
    return efficiency, fake_rate, pt_res, eta_res, n_res
```

## Common pitfalls

- Performance on photons, electrons, and muons is not reported in detail due to DELPHES dataset limitations and parametrized tracking efficiency.
- Model was trained on an unweighted ttbar sample, causing underprediction in the high-pT tail for neutral hadrons and photons.
- No event or particle weighting was applied during evaluation, which may skew performance on rare kinematic configurations.

## Evidence (verbatim from paper)

> In Fig. 7, we see that the  $\eta$ -dependent charged hadron efficiency (true positive rate) for the MLPF model is somewhat higher than for the rule-based PF baseline, while the fake rate (false positive rate) is equivalently zero, as the DELPHES simulation includes no fake tracks.

## Citation

```bibtex
@misc{pata2021mlpf,
  title={MLPF: Efficient machine-learned particle-flow reconstruction using graph neural networks},
  author={Pata et al. (2021)},
  year={2021},
  note={arXiv:2101.08578}
}
```

- arXiv: 2101.08578