Study of materials structure-property relationship using machine learning

Abstract

Recently, machine learning approaches have emerged in many applications, such as autonomous driving, text recognition, and language processing, from their flexibility in data processing despite the data acquired from various formats. Thus, materials researchers pay attention to machine learning methodology to solve many conundrums behind complex knowledge and intuitions. However, measurements in many scientific areas usually generate a small-sized and sparse dataset that is difficult to be regarded as big data, which is a common requirement for machine learning applications. In order to facilitate machine learning techniques on materials subjects, understanding how to identify and process in the form of machine-understandable style is essential. In this study, we conduct segmentation of Al-Zn alloy solidification images using X-ray computed tomography separating dendrite and liquidus phase using SegNet architecture trained on the synthetically generated phase-field training dataset. Image modifications were applied to imitate statistical metrics of the target image to prepare the training images, and the influences of each modification were inspected using accuracy, intersection over union, and boundary F1 score. In addition, the optimal number of training images for SegNet architecture was verified through quantitative metrics and visual aspects of segmentations. The best network trained on synthetic datasets showed comparable scores to those trained on manually segmented images. Afterwards, we expanded the model into 3D from 2D to transfer more context along the z-direction. The best 3D network segmented the dataset with less jittering in the vertical section. Secondly, we utilize machine learning to predict the electrochemical property, which includes composition and cycled states, of Ni-Co-Mn oxide cathode materials based on images from scanning electron microscopy (SEM). Physical parameters, including primary particle size, packing density, and secondary particle size, were considered by training the EfficientNet-b7 architecture. Classification of SEM images was preceded to verify that the convolutional neural network understands the correlation between morphologies in the image and electrochemical states, including composition and cycling state. The best model classified the composition and cycled state with 99.8% accuracy, indicating that the machine learning methods can be applied to material image datasets. In order to compare the performance of the model, we surveyed domain experts with an average of 30.0% accuracy, which equals one-third of the machine learning approach. Moreover, we conducted a grad-class activation map technique to visualize the model's decision-making process. The boundary of the secondary particles positively affects the prediction results, while the body of particles and void negatively affect the prediction. Thus, the model utilizes the boundary information to infer the composition and cycling states likewise domain experts. Furthermore, the best model expanded to SEM images of electrode materials with additives. The compositional accuracy was 96.0%, implying that we can expand our methods to embrace other electrode materials for fast analysis. Lastly, we applied machine learning methodology on atomic force microscopy, especially ferroelectric domain behavior, using Piezoresponse force microscopy and Polarization-derived friction microscopy. In order to visualize the domain evolution, we need high-speed imaging of ferroelectric domain configuration through friction force derived by polarization. Various correction methods for friction force microscopy images were applied to have high-quality images. Afterwards, super-resolution using a convolutional network named SRCNN was done. The model upscaled the image with the noise if the low-resolution images contained friction noises. In order to remove the noise and increase the quality of friction images, we applied style transfer using cyclic generative adversarial models. The friction and phase images from piezoresponse force microscopy are used to define the style. The generative model understands the specific descriptors of friction and phase images converting friction images to high-quality phase images. Thus, we can build a post-processing model to visualize the ferroelectric domain without noise if we have sufficient training data. With machine learning-assisted friction imaging, we can overcome the scanning limit of piezoresponse force microscopy that utilizes a lock-in-amplifier to calculate the signals into amplitude and phase based on input electrical signals. Various materials datasets are used for analysis to correlate the property and structures using adequate pre- and post-processing followed by machine learning approaches. Specifically, separating the meaningful information from acquired images having complicated features using machine learning can be done to secure the relationship between structure and properties. Furthermore, we can directly find the interconnection between the structure and properties faster. By applying machine learning methodology to materials discovery, we expect that we can efficiently search and evaluate new materials.