Convolutional Neural Network Based Characterization of Atomic Diffusion from Spectroscopic Imaging

In this paper, we explore the opportunities of using deep learning methods to characterize stochastic processes such as atomic diffusion. Here, we use a convolutional neural network (CNN) to characterize the extent of diffusion in a small, narrow trench. We use diffusion images to train the CNN on the diffusion profiles and categorize them into two classes. The probability of this categorization is indicative of the amount of atomic diffusion. We further explore the effects of CNN complexity (number of nodes) and different flatten functions on the loss function and the accuracy of the CNN model. We also explore different NN architectures to understand the effects of multiple hidden layers on the training and validation loss/accuracy