Feature Fusion Deep Learning Model for Defects Prediction in Crystal Structures

Abstract

Detection of defective crystal structures can help in refute such defective structures to decrease industrial defects. In our research, we are concerned with Silicon nitride crystals. There are four types of crystal structure classes, namely no-defect structures, pristine crystal structures, defective random displacement crystal structures, and defective 25% vacancies crystal structures. This paper proposes a deep learning model to detect the four types of crystal structures with high accuracy and precision. The proposed model consists of both classification and regression models with a new loss function definition. After training both models, the features extracted are fused and utilized as an input to a perceptron classifier to identify the four types of crystal structures. A novel dense neural network (DNN) is proposed with a multitasking tactic. The developed multitask tactic is validated using a dataset of 16,000 crystal structures, with 30% highly defective crystals. Crystal structure images are captured under cobalt blue light. The multitask DNN model achieves an accuracy and precision of 97% and 96% respectively. Also, the average area under the curve (AUC) is 0.96 on average, which outperforms existing detection methods for crystal structures. The experiments depict the computational time comparison of a single training epoch of our model versus state-of-the-art models. the training computational time is performed using crystal structures diffraction image database of twelve image batches. It can be realized that the prediction computational time of our multitasking model is the least time of 21 s.