AI Can Identify And Analyze Defects In Nuclear Reactors

  • Researchers applied machine learning-based model to electron microscopy to analyze defects in materials used in nuclear applications. 
  • It can identify and locate different types of defects, including stacking fault tetrahedral, grain boundaries, dislocation lines and loops, with superior accuracy. 

Developing materials that can withstand the force of nuclear power is crucial for maintaining the integrity of nuclear reactors. Even a small defect in materials could greatly affect the durability of nuclear reactor facilities and designs.

However, examining each material manually is tedious and time-consuming process. Also, manual inspection is prone to error and doesn’t scale well.

Therefore, researchers at the Oak Ridge National Laboratory and the University of Wisconsin-Madison have developed a machine learning model that can identify and analyze material damage for nuclear applications, with superior accuracy and in a very short amount of time.

They focused on electron microscopy to analyze the locations and sizes of defects of materials under irradiation. The model can be configured to access a wide range of features in microscopic pictures, from detecting line dislocations to counting nanoparticles.

Issues With Identifying Defects Manually

The manual identification process is usually time-consuming. Sometimes it takes up to an hour to investigate a single image, and thus accessing large and complex datasets can easily take days or even weeks. And since humans are prone to error, one can falsely identify or miss a fault.

Manual identification also lacks reproducibility, consistency and doesn’t scale well. Results could be affected by the human bias and training, making it quite hard to compare the final outcomes across people or establish absolute behavior.

Electron microscopes, on the other hand, can capture thousands of images per second, thanks to advances in high-speed detectors. Automated analysis techniques leverage these emerging data generation capabilities, and modern image analysis tools to detect defects could decrease the analysis-time to almost zero., yielding precise, unbiased and consistent results, which can be linearly scaled with available computational resources.

Applying Machine Learning To Electron Microscopy

Previous image recognition tools used binary feature histogram, scale-invariant feature transform, and oriented-gradient histogram to extract and enter the image features into a classifier.

Common types of defects shown in scanning transmission electron microscopy | Courtesy of researchers 

Now researchers have begun introducing machine learning techniques into the field of electron microscopy for analyzing image more efficiently. This includes detection, segmentation, classification, statistical representation, local transformation at atomic level, identification of electron diffraction patterns, and more.

The technique developed in this study can be categorized into 3 stages:

  1. A detection component with cascade object detector.
  2. A screening component with convolutional neural network.
  3. An analysis component with two local image analysis approach: one is watershed flood algorithm to locate defects and another is region property analysis to extract the sizes of those defects.

Reference: NPJ Computational Materials | doi:10.1038/s41524-018-0093-8 | WISC

The model is trained on more than 60,000 images, using a single NVIDIA GeForce GTX 1070 GPU with Matlab deep learning framework powered by CuDNN. The network successfully identified and classified 86% of dislocation errors in alloys, whereas humans detected 80% of the defects.

Read: New Algorithm Can Speed Up Material Discovery By 1000 Times

The training dataset had a limited number of microscopic pictures: one could enhance the performance of the neural network by introducing more well-annotated pictures. The authors believe that the model can be made more reliable and robust to systematic issue if supplied with more micrographic images of different materials with various contrast, brightness, focus and defect sizes.

Written by
Varun Kumar

Varun Kumar is a professional technology and business research analyst with over 10 years of experience. He primarily focuses on software technologies, business strategies, competitive analysis, and market trends.

Varun received a Master's degree in computer science from GGSIPU University. To find out about his latest projects, feel free to email him at [email protected]

View all articles
Leave a reply