Speaker
Description
In the Swedish nuclear industry, characterization of legacy radioactive waste, produced before contemporary standards, remains a significant challenge. These waste packages often involve added complications, including concrete shielding, varying origins and lacking documentation. Because of this, traditional approaches to passive characterization may struggle.
We introduce a custom convolutional neural network approach designed to estimate the locations of radioactive soures within concrete-lines waste containers. Because of the complexities of this type of waste, experimental training data is difficult to obtain. The network is therefore trained on synthetic data produced via Monte Carlo-simulations to learn to map complex signal patterns to three-dimensional spatial coordinates for bounding boxes containing the sources.
The neural network model is designed to handle varying detector geometries and emission spectra, provided that it has been trained on the corresponding type of data, as it is independent of sensor position information. Preliminary testing on multi-source scenarios demonstrates the model's ability to detect both the quantity and location of the sources within the waste containter volume. These findings suggest that automated machine learning methods can serve as a valuable tool for characterization in complex waste scenarios.