Feasibility study of machine learning-based discrimination of α, β and γ particles from grayscale radiation images

Abstract

In the framework of the decommissioning of the Fukushima Daiichi Nuclear Power Station, the Japan Atomic Energy Agency analyses and classifies a variety of radioactive samples. The objective is to simplify the sample characterization process by developing multipurpose analysis tools that quickly produce results for different types of samples while reducing labour.

The development of an analytical device has been started, based on the MiniPIX TPX standard detector, a hybrid semiconductor pixelated radiation detector. The final aim is to build a fast-mapping device that generates 2D activity maps, distinguishing between α, β, and γ radiation, and includes simple local γ spectrometry for highly contaminated samples.

This study explores the use of supervised machine learning models to classify individual particle interactions based on nine morphological and intensity-based features extracted from clusters recorded during exposure to pure α-, β-, and γ-emitting sources (⁶⁰Co, ⁹⁰Sr/⁹⁰Y, ¹³⁷Cs, ²⁴¹Am). The best performing models demonstrated the best balance of accuracy (∼79–80 %) and processing speed (microseconds per frame). α particles were reliably identified with 100 % accuracy.

However, a detailed energy-binned analysis revealed that β and γ interactions are not distinguishable when they produce similar energy depositions and cluster morphologies. In particular, low-energy β events were frequently misclassified as γ and vice versa at high energies due to overlapping track structures.

These findings highlight the limitations of shape-based classification under real-world conditions where energy spectra overlap and the current application could only be the identification and mapping of α vs. β/γ events.