Deep Learning-Based Isotope Identification for Radiological Crime Scene Investigations Using Convolutional Neural Networks

Abstract

We address on-scene nuclear forensics with a portable NaI 3 inch detector by training a convolutional neural network on spectra synthesized from Geant4 event level pools with bootstrap resampling, mixed with measured background, and regularized via global channel-shift augmentation to mirror handheld calibration drift. The model performs single isotope identification across 33 radionuclides. On held out synthetic tests spanning low to high signal fractions and 5k–100k signal counts, it reaches 91.5% overall accuracy, rising from ~75% at the lowest signal fractions and stabilizing near 94% in mid/high fractions. Real data validation with a NaI 3 inch system using six isotopes, 1000 spectra per isotope created by random resampling to 5k–100k events from the measured spectra, yields 84.5% overall accuracy, with Am-241 and Co-60 at 100%, Eu-152 at 99%, Cs-137 at 83.7% , Ba-133 at 85.5%, and Pu-239 at 38.9% due to systematic confusion with Pu-240, which rises to 97.5% when reported as the combined class Pu-239/Pu-240. The results demonstrate a practical, reproducible pipeline that works with simple fieldable detectors and provides a clear path toward mixture detection and unmixing.