The Effect Of Input Features On Advanced Gamma Ray Spectra Analysis

Gamma-ray spectroscopy is a powerful, non-destructive, technique that can be used to obtain isotopic information from samples of interest. However, features associated with certain isotopes can be challenging to extract in complex detection environments. Various algorithms, and analysis techniques, are utilized to obtain information from gamma-ray spectra, but often without generalizable applicability to different, or changing, measurement environments. An advanced analysis method that has proven successful at extracting information and patterns from complex data is machine learning. Machine learning models have enabled significant advancements in the medical, financial, and robotics sectors, and recent applications to the field of nuclear science and engineering, specifically with regards to radiation detection, have shown promise for improved radioisotope identification and radiation imaging performance. However, one of the challenges associated with the application of machine learning algorithms to radiation detection data is the effect of variations in measurement scenarios that are not well captured during the training process. These variations can take the form of changing background, relative changes in isotope intensity, or the presence of isotopes which have not been included during the training phase. To determine the effect of these variations on the spectral analysis capabilities of different machine learning algorithms, the impact of various spectral features is examined. Well-controlled data sets were produced using the Monte Carlo based radiation transport code MCNP to simulate spectra for different isotopes, backgrounds, and detectors. Several spectral feature, such as full spectra, peak height, and peak-to-Compton ratios, were used as inputs for different supervised machine learning algorithms, including K-Nearest Neighbor, Support Vector Machines, Logistic Regression, and Naïve Bayes. Presented here is the ability of the different algorithms to extract isotopic information under simulated scenarios as a function of varying input features. The algorithms used for this effort were selected because they belong to several different classes, providing a comparison of algorithm performance, and a basis for future algorithm development.