This application involves the development of a CNN-based Machine Learning Model for image classification of barrier coatings microstructures in elastic stress contours, with the goal of accelerating thermo-mechanical fracture tests and reducing the runtime of finite element simulations. Built using the Kaggle Jupyter IDE and the NVIDIA T4 x2 Tensor Core GPU.
Overview:
Convolutional NN Features:
- Initial inputs of 150x150 images with 4 channels
- Batch size = 32, Weight decay = 1e-4, CrossEntropy loss, Adam optimizer
- Additional hyperparameter tuning:
- 5-Fold cross validation
- Batch normalization
- PyTorch tensors for xData and yData
- Learning rate scheduler
Accuracy Performance Metrics- Training Accuracy: 99.59%, Test Accuracy: 93.10%
Loss Performance Metrics- Training Loss: 0.022, Test Loss: 0.235
Other Classification Metrics- Average Precision: 0.887, Average Recall: 0.890, Average F1 Score: 0.898
ROC Curve and AUC-
