Prediction of material defects and phase formation using deep learning and transfer learning

Machine learning has been successfully employed in computer vision, speech processing, and natural language processing. However, when machine learning is applied to materials study, many challenges remain, and they include small datasets, tricky manual feature (descriptor) engineering, isolated models and poor interpretability. In this project, possible solutions to these challenges have been explored using deep learning and transfer learning.

For the challenge of small datasets, fully connected deep neural network and tree-based models were used to predict solidification cracking susceptibility of stainless steels with a dataset of 487 samples. It is found that deep neural network with pre-training and fine-tuning improves prediction accuracy, and tree-based models reveal the relative importance of input variables.

To overcome the challenge of tricky manual feature engineering in predicting phase formation in inorganic substances and compounds properties, I proposed a general and transferable deep learning framework as follows: (1) mapping raw data to pseudo images with periodic table structure, (2) automatically extracting features through convolutional neural networks, (3) transferring knowledge by sharing features extractors between models. The proposed deep learning models outperformed previous models in predicting glass-forming ability using a medium dataset of 16k samples and compounds properties using a big dataset of 228k samples. The developed transfer learning model for multi-principal element alloys can distinguish five phases (BCC, FCC, HCP, amorphous, mixture) with high scores (0.94) in a small dataset of 345 samples. The transfer learning model for phase prototypes can discriminate 170 phase prototypes with an accuracy of 0.9 in a dataset of 17k inorganic substances. Periodic table knowledge embedded in data representations and knowledge shared between models is beneficial for tasks with small datasets.