code_public/radiacode
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix: mask channels below 30 keV in inference and training to prevent misidentification

Browse Source 
Below ~30 keV the detector signal is dominated by X-ray fluorescence (L-shell)
and artifacts not modelled in training data. This spurious low-energy continuum
caused the model to misidentify Am-241 as Th-232/U-235. Masking channels <30 keV
before inference fixes Am-241 detection from 2% to 99%. Same masking applied in
the synthetic spectrum generator for consistent retraining.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Jacquin Antoine
2026-05-21 21:55:34 +02:00
parent 00418d35bc
commit 091d7d9eb8
2 changed files with 91 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
train/vega_ml/synthetic_spectra/generator.py
View File

@ -307,6 +307,12 @@ class SpectrumGenerator:
# Subtract and clip — same as inference: net = clip(rate - bg_rate, 0, inf)
spectrum = np.maximum(spectrum - bg_spectrum2, 0)
# Mask channels below ~30 keV — below this energy the detector signal is
# dominated by X-ray fluorescence and artefacts not modelled in training.
min_channel = max(0, int((30.0 - self.detector_config.calibration_offset_kev)
/ self.detector_config.calibration_slope_kev))
spectrum[:min_channel] = 0
# Normalize if requested
if config.normalize:
spectrum = normalize_spectrum(spectrum, config.normalization_method)