Oct 20, 2023
Contact: Sarah Day
Summary
Homecoming Special Event: Victoria Gerardi '17 BS, '18 MS
Full Description
Title: Using Machine Learning for Defect Characterization
Abstract: This talk presents a methodology that is under development to analyze large X-ray image datasets for anomaly and/or defect detection using machine learning techniques. The characterization of anomalies and/or defects can be identified through the performance accuracy of either image classification (supervised learning - convolutional neural
networks) or anomaly detection (unsupervised learning - autoencoders) models. Each learning technique has unique hyperparameters and design architectures to aid in creating a robust model to predict against X-ray images of varying orientations, brightness and contrast. This method would be a strong complement to the traditional suite of energetic material/component characterization tests, particularly for melt-pour explosives, performance-related design intent, safety, and/or performance-related defect detection. For safety or performance-related defect detection, the methodology enables baselining defects as a feedback loop in the development of new subscale tests and physics-based models to better understand and predict energetic failure modes, a capability under development at DEVCOM Armament Center called Energetic Defect Characterization (EDC).
