A COMPARISON OF MATHEMATICAL ALGORITHM AND DEEP NEURAL NETWORKS FOR RADIOACTIVE PARTICLE TRACKING TECHNIQUE

Resumo

Radioactive particle tracking (RPT) is a nuclear technique used to track a sealed radioactive particle within a volume, commonly applied in hydrodynamic flow studies. This report compares the effectiveness of traditional mathematical algorithms and deep learning models in reconstructing the trajectory of a radioactive particle. The traditional algorithm involves solving a minimization problem between measured events and a calibration dataset, implemented in C/C++. A six-layer deep neural network (DNN) was also developed, with hyperparameters optimized using Bayesian methods. The study evaluates the impact of the calibration dataset size on accuracy for both approaches. The simulation model includes a concrete mixer, six NaI(Tl) detectors, and a ¹³⁷Cs radioactive particle, with measurement geometry and datasets generated using the MCNPX code. Results demonstrate the superior accuracy of the DNN in the RPT system. This report highlights the potential of deep learning models in improving the accuracy of RPT compared to traditional mathematical algorithms.