Strengthening Nuclear Security with ML, Full-Spectrum 137Cs Burial Depth Estimation

Abstract

The non-intrusive characterization of buried radioactive sources is a critical capability for thwarting illicit trafficking, mitigating orphan‑source hazards, and safeguarding civilian populations against radiological threats. Depth estimation, in particular, enables rapid threat assessment and informed countermeasure deployment following incidents such as transnational uranium diversion or the loss of medical and industrial sources. In this feasibility study, we demonstrate a machine learning approach to estimate the burial depth of a 137Cs point source in dry sand over the range of 5–95 cm. Our method employs gradient-boosted decision trees trained on simulated full gamma-ray spectra partitioned into 1024 energy bins, thereby exploiting subtle variations across both the Compton continuum and multiple photopeaks. After hyperparameter tuning, the model achieved an average depth‐estimation standard deviation of 5 cm across the full depth range. By leveraging the entire spectral profile rather than isolated peak ratios, this algorithm delivers enhanced accuracy and robustness in heterogeneous field conditions. The results validate the potential of full-spectrum, gradient boosted models as field‑deployable tools for rapid subsurface threat localization, reinforcing layers of nuclear security and environmental monitoring efforts worldwide.