Project Introduction: Material properties are closely related to material microstructure, with many journal/conference papers explicitly stating that structure directly determines performance. Using deep learning methods to analyze material microstructure images to build regression models that predict material properties is an excellent approach. However, many current works simply transfer methods that perform well in other visual tasks directly to the material image domain to predict material properties, such as directly transferring ResNet, AlexNet, etc., which perform well on COCO and ImageNet, to material images, replacing the final classification network with a network that outputs one category (performance value), and then training it to predict material performance. However, in materials science papers, it's also clear that material performance is not only limited to specific structures in material images. In fact, global features are as important as local features. Local features characterize the material's texture, grain boundary density, and porosity in detail, while global features consider long-range dependencies between different features and can discover interaction relationships between different positions in materials, considering more comprehensive information to establish a more complete mapping relationship. The network built by this program considers both local network branches and global network branches, extracts local features and global features of material images respectively, and proposes a specialized feature fusion module for fine-grained fusion processing of features, ultimately accurately predicting performance.

Software Introduction: The nanostructures on ceramic surfaces are extremely complex and variable. Slight differences in material composition can lead to significant changes in material images. This research platform was developed for segmentation and recognition of various regions in ceramics (HfB2−B4C), as shown in Figure 1. By finding region boundaries, each region can be identified, and then further identification of subregions within each region can be performed.

Software Introduction: To study the morphology of DNA robots, researchers use atomic force microscopes to capture AFM images of the robots. Due to the high noise in AFM images and the large-scale overlap of the observed DNA robots, classification and recognition of DNA nanorobots are needed. DNA robots in AFM images include three different DNA morphologies: parallel, anti-parallel, and crossed. Because they are made of DNA macromolecules and photographed in liquid, one of the difficulties of the recognition task is the diverse and flexible shapes of these nanorobots in AFM images. Although nanoscience and technology have developed rapidly in many fields, it is still difficult to obtain enough samples of nano-objects due to high costs, which hinders the development of deep learning methods in the field of materials. Based on these situations, a DNA robot recognition software based on transfer learning and convolutional neural networks was designed.

Software Introduction: Electron backscatter patterns can reflect crystal structure and orientation information through analysis, and their accuracy depends on the determination of pattern position and pattern axis lines. The project studied a new method for processing Kikuchi band patterns, which can quickly and accurately label the Kikuchi bands, their axial lines, and intersection points in the image. First, edge contours are extracted using the Canny operator, and then Radon transform line detection is used on the results to obtain the two edges of each Kikuchi band and merge them to get the center line. Subsequently, the probabilistic Hough transform method is used to detect edge line segments in the original image. After using a bidirectional filtering algorithm to select and pair the resulting center lines and line segments, a hyperbola is used to fit the edges of the Kikuchi band. Figures 5 and 6 show the line and line segment detection results of the Kikuchi band pattern, and Figure 7 shows the pairing results of two Kikuchi band patterns. The final detection result of the Kikuchi band boundary is shown in Figure 8.

Software Introduction: This software accomplishes automatic extraction of lamella contours based on mathematical morphology and calculation of spreading morphology parameters. It uses maximum inter-class variance method to determine binary segmentation threshold, applies mean filtering and morphological operations for image denoising, and ensures connectivity of individual lamellae. Lamella edge information is obtained through contour extraction, and finally, lamella solidity parameters are calculated based on the extracted contours. This software can accurately identify lamella contours and has strong anti-noise interference ability, effectively improving the research efficiency of scientific researchers.

Software Introduction: The development of new materials is an important driving force for the advancement of materials science, and materials genome engineering is a frontier interdisciplinary field in materials science that can shorten the development cycle of new materials and reduce time and human resource costs. Segmentation and recognition of material image microstructures can provide a data foundation for database construction in materials genome engineering. This method proposes a material image segmentation approach based on complex texture feature fusion, selecting and improving FCN as the basic network to achieve precise segmentation of material images with similar textures across phases. The network is divided into encoding and decoding stages, using VGG16 (VGG block) as the backbone network. In the decoding stage, cascaded feature fusion modules are used to merge semantic information from both high-level and low-level feature maps; then the fused feature maps are fed into a multi-scale learning module (Multi-scale block) to further extract texture information. In the decoding stage, attention mechanism modules (Attention block) are applied to each layer of restored feature maps to add weights, where channel attention mechanism preserves important feature maps and spatial attention mechanism preserves key texture information in important feature maps. At the same time, for the problem of unbalanced data distribution in material images, this chapter selects and improves Dice loss to optimize segmentation results.

Software Introduction: Due to atomic motion, crystal structures change, leading to plastic deformation. The plastic deformation of crystalline materials is closely related to material properties, and studying material deformation mechanisms is extremely important for analyzing material performance and understanding deformation mechanisms. However, crystal structure recognition often faces the problem of insufficient sample data, and some existing methods cannot identify all crystal structures. In view of the small sample problem in crystal structures and the deficiencies of existing methods, this study proposes a crystal structure recognition method based on machine learning.

Software Introduction: When preparing materials, some precipitated precipitates will affect the performance of the material, and the size and distribution of precipitates have a significant impact on performance. For example, duplex stainless steel has good mechanical properties and corrosion resistance, which are attributed to the formation of ferrite and austenite, but secondary phases such as σ phase are easily produced during the preparation process. For ease of analysis, images can be collected and processed during the preparation process. Based on the good performance of deep learning and its successful applications in materials research, this program provides multiple segmentation networks and models trained on some datasets to help users more conveniently process material images with deep learning algorithms. The interphase energy is calculated according to the Ostwald ripening mechanism, based on the radius of precipitates at each moment.