Calorimetry Reconstruction Eval

calorimetry-reconstruction-eval

Calorimetry with Deep Learning: Particle Simulation and Reconstruction for Collider Physics — Belayneh et al. (2019) (arXiv:1912.06794, 2019)

What this evaluates

Evaluates deep learning models for high-energy physics calorimetry tasks, specifically particle shower generation and particle reconstruction (identification and energy regression) using simulated detector data.

Datasets

• LCD Calorimeter Dataset (GEN & REC) — total ?; splits: GEN (-1), REC (-1)

Metrics

• accuracy (primary) — range: percent
  • Classification accuracy for particle type identification (e, γ, π, π⁰). Exact formula not specified in provided text.
• regression error — range: other
  • Error metric for regressed particle energy against GEANT4 ground truth. Exact formula not specified in provided text.

Input / output format

Input: 3D arrays of energy deposits in calorimeter cells. GEN input: 51×51×25 ECAL window. REC input: 25×25×25 ECAL slice concatenated with 11×11×60 HCAL slice.

Output: GEN: 3D voxel image representing the calorimeter shower. REC: Discrete particle type label and continuous regressed energy value.

Scoring recipe

def score(predictions, gold):
    correct = sum(1 for p, g in zip(predictions, gold) if p[0] == g[0])
    accuracy = correct / len(gold)
    energy_errors = [(p[1] - g[1])**2 for p, g in zip(predictions, gold)]
    mse = sum(energy_errors) / len(energy_errors)
    return {'accuracy': accuracy, 'mse': mse}

Common pitfalls

• Dataset is entirely GEANT4-simulated, lacking real detector noise or calibration effects.
• Task-specific filtering (HCAL/ECAL ratio < 0.1 for pions, opening angle < 0.01 rad for π⁰) removes easy cases, creating a biased evaluation subset.
• Fixed window sizes may truncate shower tails for high-energy particles, affecting generalization.

Evidence (verbatim from paper)

When training classification models on these data, a negligible accuracy increase was observed when moving to larger windows, as described in Appendix[A]... We also compute a set of expert features, as described in Ref.[NIPS]. These features are used to train alternative benchmark algorithms (see Appendices[C] and[D])...

Citation

@misc{belayneh2019calorimetry,
  title={Calorimetry with Deep Learning: Particle Simulation and Reconstruction for Collider Physics},
  author={Belayneh et al. (2019)},
  year={2019},
  note={arXiv:1912.06794}
}

• arXiv: 1912.06794