| About this Abstract |
| Meeting |
MS&T21: Materials Science & Technology
|
| Symposium
|
2021 Undergraduate Student Poster Contest
|
| Presentation Title |
Machine Learning Approaches to Predict Properties from Microstructure Images in Ceramic-Metal Composites |
| Author(s) |
Hugh B. Smith, William Huddleston, Laura Bruckman, Alp Sehirlioglu |
| On-Site Speaker (Planned) |
Hugh B. Smith |
| Abstract Scope |
Electrical conductivities of composites of Li4Ti5O12 anode and Ni current collector particles for structural battery applications were predicted from SEM microstructure images. Further, microstructural features contributing the most to conductivity for different samples could be identified. Principal component analysis (PCA) was performed on voronoi, nearest neighbor, size, and skeleton distributions of the microstructural features, and logistic regression and linear discriminant analysis models were fit to the scores of the principal components (PCs) to classify the images as low or high conductivity. Accuracies exceeded 87%, and the most important PCs were identified. Convolutional neural networks (CNN) were then used for classification and led to accuracies above 98%. Class activation maps were created from the CNN models for testing images. These highlight which microstructural features influenced conductivity predictions. Regression was attempted on the PCs using multiple linear regression and general additive models with little success. Regression attempts with CNN were unstable due to the sparsity of the data set but yielded responses that were usually 4x off the observed responses on average with values for r2 usually exceeding 0.7. |