Towards Green and Safe Nuclear Energy Via Statistical Microscopy
Presentation Type
Poster
Abstract
This project aims to advance the rate of material science study by automating one highly time consuming task. Specifically, we use machine learning to segment high resolution images of alloys that are used in nuclear reactors. The images are generated by an electron transmission microscope and depict irradiation defects of the alloys, including dislocation lines, dislocation loops, voids and precipitates. Traditionally, material scientists have to manually label the defects in order to better understand how radiation affects materials and design longer lasting alloys as well as make more reliable lifetime estimates.
While it takes a human ~13 hours to label a single image, our system can produce labels in a fraction of a second. Our model uses a convolutional neural networks trained to make pixel-level predictions about defects. In preliminary results, we obtain area under the receiver operating characteristic curve of up to 0.82 for dislocation lines, 0.81 for dislocation loops, 0.93 for voids and 0.80 for precipitates. Work to improve the model is ongoing.
Keywords: Machine Learning, Deep Learning, Material Science, Radiation, Nuclear Energy, Statistical Microscopy, Convolutional Neural Networks
Start Date
10-5-2018 12:00 PM
End Date
10-5-2018 2:00 PM
Genre/Form
posters
Subjects - Topical (LCSH)
Nuclear energy--Research; Nuclear energy--Environmental aspects; Transmission electron microscopes; Machine learning; Materials science
Type
Event
Format
application/pdf
Language
English
Towards Green and Safe Nuclear Energy Via Statistical Microscopy
This project aims to advance the rate of material science study by automating one highly time consuming task. Specifically, we use machine learning to segment high resolution images of alloys that are used in nuclear reactors. The images are generated by an electron transmission microscope and depict irradiation defects of the alloys, including dislocation lines, dislocation loops, voids and precipitates. Traditionally, material scientists have to manually label the defects in order to better understand how radiation affects materials and design longer lasting alloys as well as make more reliable lifetime estimates.
While it takes a human ~13 hours to label a single image, our system can produce labels in a fraction of a second. Our model uses a convolutional neural networks trained to make pixel-level predictions about defects. In preliminary results, we obtain area under the receiver operating characteristic curve of up to 0.82 for dislocation lines, 0.81 for dislocation loops, 0.93 for voids and 0.80 for precipitates. Work to improve the model is ongoing.
Keywords: Machine Learning, Deep Learning, Material Science, Radiation, Nuclear Energy, Statistical Microscopy, Convolutional Neural Networks