InDEx · C₂H₂F₄ Superheated Liquid Detector · Saha Institute of Nuclear Physics

Telling a neutron's acoustic signature
from a gamma-ray's.

Every bubble nucleation in the detector fires an acoustic pulse. Neutrons and gamma-rays leave near-identical amplitudes — but their short-time Fourier spectra don't lie. This dashboard runs the full pipeline end-to-end on the actual 200-event dataset: pulse localization, 512-D spectral feature extraction, and two linear classifiers trained from scratch.

Neutron —
Nuclear recoil~20–40 kHz shoulder
Gamma —
Electron scatterEnergy near DC
Fs = 1.000 MSa/s · Δt = 1 µs · 20 ms window · highlighted band = localized pulse · hover a card to replay
200Waveform events
100 / 100Neutron / Gamma
512-DSTFT feature vector
7:2:1Train · Val · Test split
30,000SGD training epochs
01

Why bother discriminating

◎

The dark matter background problem

  • ~85% of the universe's matter has never been directly observed — dark matter interacts too weakly with ordinary matter to detect easily.
  • Rare-event searches need detectors sensitive enough to catch a single nuclear recoil, which also makes them sensitive to everything else nearby.
  • The central challenge: separating a genuine dark-matter-like recoil from ordinary radioactive background.
⚛

Neutrons as a dark-matter stand-in

  • Dark matter is expected to scatter off atomic nuclei — a nuclear recoil.
  • Neutrons produce the same kind of nuclear recoil, making them the standard laboratory proxy for dark matter signals.
  • Gamma-rays instead scatter off electrons, producing a different acoustic signature that forms the main background.
  • Reliable neutron/gamma discrimination is therefore a direct proxy for validating dark-matter search sensitivity.
02

From raw pulse to prediction

03

Spectral separation between classes

Mean STFT feature value at each frequency bin, averaged across every event of each class. Gamma events concentrate almost all their spectral energy at DC; neutron recoils show a distinct ~20–40 kHz shoulder. The pattern holds independently in the train and test partitions — the separation isn't an artifact of the split.

MEAN FEATURE VALUE vs FREQUENCYTRAIN SET
Neutron Gamma
04

Live classifier

Drop in a raw .lvm waveform and this page runs the exact pipeline above — normalization, pulse localization, STFT, 512-D feature extraction — entirely in your browser, then scores it with the actual trained LR and SVM weights. Nothing is uploaded anywhere.

⤓
Click to choose a .lvm file, or drag one here — try any file from the Neutron/ or Gamma/ folder.
05

Sample event library

Pick any of 20 sample events (10 neutron, 10 gamma) pulled from across the train/val/test split. Each panel is computed from that event's actual waveform — pulse localization, STFT spectrogram, and both trained classifiers' live predictions.

RAW WAVEFORMcrop —
SHORT-TIME ENERGYθ = 0.15
STFT SPECTROGRAM (cropped window)0–500 kHz
Logistic Regression
Linear SVM
06

Model performance

TRAINING LOSS vs ITERATION30,000 epochs · SGD, invscaling lr
LR — log-loss SVM — hinge-loss

Logistic Regression — confusion matrix (test set)

Linear SVM — confusion matrix (test set)

Metrics summary — held-out test set (n = )