University of Edinburgh
boosted di-boson resonance search

The analysis

Boosted di-boson resonance search with mass-decorrelated jet tagging

The search targets a heavy resonance X decaying to a pair of electroweak bosons (H → ZV, W′ → WZ, Z′ → ZH) in the boosted regime, where the decay products of each boson merge into a single large-radius fat jet. The analysis reconstructs two fat jets and searches for a bump in the reconstructed di-jet invariant mass m_jj above a smoothly-falling QCD background.

The mass sculpting problem is acute in this topology. A boosted-object tagger trained to separate W/Z/H jets from QCD will learn the jet mass as a discriminating feature: signal jets have a characteristic mass around 80–125 GeV; QCD jets have a steeply falling mass distribution. Any cut on the tagger score preferentially removes low-mass QCD background, carving a bump-shaped enhancement into the m_jj spectrum at the signal hypothesis. A bump-hunt that treats the tagger as mass-decorrelated will attribute this enhancement to a new resonance. The significance is fake.

The existing analysis pipeline uses a neural network tagger trained on eight high-level kinematic variables per fat jet: lep1_pt, lep2_pt, fatjet_pt, fatjet_eta, fatjet_D2, Zll_mass, Zll_pt, MET. Vikshep replaces this tagger with constituent-level wavelet scattering features plus the DisCo mass-decorrelation penalty. The physics reach is measured by two numbers against the same Wilks Δχ² significance pipeline the analysis already uses.

The data

Geant4-simulated ATLAS samples

The pilot runs on Geant4-simulated ATLAS detector response with full detector geometry. Signal samples cover gg → H → ZV at four resonance mass points (700, 1000, 1500, 2000 GeV). Background samples include diboson production (WW, WZ, ZZ), Z+jets (with up to four additional partons), and semi-leptonic top-quark pair production.

Each fat jet is reconstructed with the anti-k_T algorithm at R=1.0 and trimmed with f_cut=0.05. Constituent four-vectors are read directly from the ROOT TTree output of the ATLAS analysis framework. The jet image rasterisation is handled by the Vikshep ingest step at 64×64 bins in (η, φ) centred on the fat jet axis.

The benchmark

Two numbers. Apples-to-apples.

The benchmark is computed against the analysis’s own Wilks Δχ² significance pipeline — the same code the collaboration uses to quote significance in the paper. No new significance metric, no cherry-picked working point. Two numbers are reported:

Success criterion: Vikshep achieves equal or higher Δσ at strictly lower ΔJSD. A higher-significance result that sculpts more than the NN baseline is not a success — it would trade a physics guarantee for a statistical number. The goal is to demonstrate that the two objectives are not in tension: you can have both.

Status & results

Currently in progress.

Reproduce

The full harness is in the repo.

The pilot analysis harness — data loading, scattering configuration, DisCo training loop, and Wilks significance pipeline — lives in Vikshep/pilot/ in the GitHub repository. To reproduce:

1. Clone the repo: git clone https://github.com/samvardhan03/Vikshep
2. Install dependencies: pip install -e backend/ingest
3. Obtain the Geant4 samples (see pilot/README.md for access instructions)
4. Run the harness: bun run pilot/run.ts --config pilot/uoe-diboson.yaml
5. Results are written to pilot/results/ as JSON + ROOT files

Geant4 Direct Interface

Run this on your own data.

No hosted compute required. Clone the repo, install the ingest package, and go from Geant4 CSV export to features and calibration plots locally — under 30 seconds on a fresh machine, no GPU.

Install (repo-local; not yet on PyPI)

git clone https://github.com/samvardhan03/Vikshep
cd Vikshep
pip install -e backend/ingest

Four-command pipeline

# 1. Ingest your Geant4 CSV
vikshep-ingest g4 your_output.csv --schema komal_v1

# 2. Calibrate detector response
vikshep-recipe calibrate \
  --features manifest.json --target energy_mean

# 3. Tag with DisCo decorrelation
vikshep-recipe tag \
  --features manifest.json \
  --label is_signal --protect mass --lambda 1.0

# 4. Run the benchmark harness
python -m bench.run \
  --manifest manifest.json \
  --label is_signal --protect mass

The komal_v1 schema expects per-hit CSV rows with event_id, layer, phi, theta, momentum[, energy]. Use --schema generic --column-map for other ntuple exports. A working sample CSV and step-by-step walkthrough live in examples/g4_quickstart/.

For research groups

Run Vikshep on your own Geant4 simulation.

We are actively looking for a second pilot partner — a research group with Geant4 simulation output and an existing analysis pipeline. We will integrate Vikshep into your analysis, validate against your existing significance benchmark, and deliver the two numbers that matter: Δσ and ΔJSD. No hosted compute required; the tool runs locally on your data, in your facility.

We ask for one thing: permission to cite the results (anonymised if needed) and to share the benchmark numbers publicly. If the results are not better than your existing pipeline on both metrics, we will say so plainly.

Cite

@misc{vikshep2025,
  title        = {Vikshep: Deterministic Wavelet-Scattering Features
                  for Scientific Compute},
  author       = {Singh, Samvardhan and Mishra, Yash and Papanwar, Komal},
  year         = {2025},
  howpublished = {\url{https://github.com/samvardhan03/Vikshep}},
  note         = {AGPL-3.0. Engine ships as binaries under separate terms.}
}

Contract statement: participation in a pilot is not a commercial engagement. No fees. No data leaves your facility. The ingest step runs locally; only the benchmark numbers (scalars) are shared.

Acknowledgements

Built with.