Determination of radioisotopes in gamma-ray spectroscopy using abductive machine learning

An algorithmic approach has been adopted for many years for identifying and quantifying radioisotopes in high-resolution gamma-ray spectra. Complexity of the technique, particularly when used with lower resolution detectors, warrants looking for machine-learning alternatives where intensive computations are required only during training, while actual sample analysis is greatly simplified. This should be advantageous in developing simple portable systems for fast online analysis of large numbers of samples, particularly in situations where accuracy can be traded off for speed and simplicity. Solutions based on neural networks have been reported in the literature. This paper describes the use of abductive networks which offer shorter training times and a simpler and more automated approach to model synthesis. The Abductory Induction Mechanism (AIM)(1) tool was used to build models for determining isotopes in both single- and multiple-isotope samples represented by spectra from an NaI (T1) detector. In spite of a 50-fold poorer resolution for the AIM spectral data, AIM results are adequate, with average errors ranging between 11.8% and 20.5% for a number of simulated multi-isotope cocktails.