Final Year Project: Predicting Charpy Impact Energy from Microstructure Data of A508 steel with Machine Learning
This study develops a machine learning-based model to predict Charpy impact energy from microstructural data of A508 steel, commonly used in nuclear reactor pressure vessels. Previous research focused on process parameters for material property prediction but overlooked microstructural data.
Feature extraction techniques such as Histogram of Oriented Gradients (HOG) and transfer learning models (VGG16 and EfficientNetB0) were compared to extract features from microstructural images. These features were combined with machine learning models like fully connected neural networks (FCNN), random forest (RF) regression, and support vector regression (SVR) to predict Charpy impact energy.
The RF regressor with EfficientNetB0 exhibited the lowest mean absolute error (MAE) of 4.5 J, significantly improving upon the baseline MAE of 12.1 J without microstructural data. This demonstrates a clear correlation between microstructural images and Charpy impact energy, highlighting the potential for machine learning in non-destructive material characterization.
Keywords: Charpy test, microstructure, A508 steel, machine learning, transfer learning, EfficientNetB0, VGG16, histogram of oriented gradients, support vector regression, random forest regression
For the full dissertation, please email me at tszyyung@gmail.com.