Feasibility study of fast nuclide identification for beta-emitting sources using learning-based models

This study investigates the feasibility of rapidly identifying beta-emitting radionuclides based on the beta spectrum data measured in air using learning-based models. Although beta particles are easily shielded, rapidly identifying beta-emitting nuclides is critical for ensuring radiological safety in radiological emergencies or during work at nuclear facilities. For example, rapid beta-emitter identification is needed during post-accident environmental surveys (e.g., radiostrontium monitoring after Fukushima) and for immediate decision-making following airborne contamination alarms from continuous air monitors in nuclear facilities. However, nuclide identification requires chemical preprocessing, high-precision detectors, or sophisticated analysis equipment because of the continuous form of the beta spectra. To address this limitation, an artificial intelligence -based approach is proposed for fast beta nuclide identification. An experimental system is setup using beta-disk sources along with an appropriate detection system and associated electronics. Two learning-based models, support vector machine (SVM) and time-series classification with transformer (TSCT), are employed to classify 15 combinations derived from ⁶⁰Co, ⁹⁰Sr/⁹⁰Y, ¹³⁷Cs, and ¹⁵²Eu. The SVM and TSCT models achieved classification accuracies of 100 and 98.0%, respectively. Once trained, both models under controlled laboratory conditions with fixed geometry and stable electronics could identify nuclide combinations within a few seconds, confirming that the proposed method significantly enhances the speed of beta nuclide identification. It is thought the present study would provide a basis to apply to the real environment with additional validation under more variable field conditions, contributing to more effective radiation protection strategies in field scenarios.