Thursday 27 March 2025
A team of researchers has made significant progress in developing a new method for analyzing microstructural evolution during irradiation of materials used in nuclear reactors. The study, published recently, uses deep learning techniques to improve the accuracy and efficiency of this analysis.
The research focuses on LiAlO2 pellets, which are used in tritium-producing burnable absorber rods (TPBARs) to produce tritium, a key component for nuclear weapons. Understanding the relationship between microstructural evolution during irradiation and tritium diffusion, retention, and release is crucial for predicting the performance of TPBARs.
Traditionally, researchers have relied on manual segmentation of images taken with scanning electron microscopy (SEM) to analyze the microstructure of these pellets. However, this process is time-consuming and prone to human error. The new method uses a deep convolutional neural network (CNN) architecture called SegNet, which is capable of learning patterns in image data and making predictions.
The researchers trained the SegNet model on a dataset of SEM images of unirradiated and irradiated LiAlO2 pellets. They used two types of loss functions: expert-weighted cross-entropy (EWCE) and weighted cross-entropy (WCE). The EWCE loss function is designed to prioritize predictions that match the expert-labeled data, while the WCE loss function is more forgiving and allows for some variation in the predictions.
The team tested their model on a range of images, including those with varying levels of irradiation. They found that the SegNet model performed well, achieving high accuracy and precision in segmenting the microstructure of the pellets. The model was able to correctly identify defects such as voids, zirconia impurities, and LiAl5O8 precipitates.
The researchers also explored the use of additional metadata, such as beam spot size and accelerating voltage, to improve the performance of their model. They found that incorporating this information into the training process improved the accuracy of the predictions.
One of the key benefits of this new method is its ability to analyze large datasets quickly and efficiently. This could allow researchers to study the microstructural evolution of LiAlO2 pellets in greater detail, which could lead to a better understanding of how they respond to irradiation.
The team’s approach has several potential applications beyond TPBARs.
Cite this article: “Deep Learning Method for Analyzing Microstructural Evolution in Nuclear Materials”, The Science Archive, 2025.
Microstructural Evolution, Deep Learning, Image Analysis, Lialo2 Pellets, Tritium-Producing Burnable Absorber Rods, Sem Images, Convolutional Neural Network, Segnet Model, Weighted Cross-Entropy Loss Function, Expert-Weighted Cross
