In this study, the average grain size was evaluated from a microstructure image using a convolutional neural network. Since the grain size in a microstructure image can be directly measured and verified in the original image, unlike the chemical composition or mechanical properties of material, it is more appropriate to validate the training results quantitatively. An analysis of microstructure images, such as grain size, can be performed manually or using image analysis software; however, it is expected that the analysis would be simpler and faster with machine learning. Microstructure images were created using a phase-field simulation, and machine learning was carried out with a convolutional neural network model. The relationship between the microstructure image and the average grain size was not judged by classification, as the goal was to have different results for each microstructure using regression. The results showed high accuracy within the training range. The average grain sizes of experimental images with explicit grain boundary were well estimated by the network. The mid-layer image was analyzed to examine how the network understood the input microstructure image. The network seemed to recognize the curvatures of the grain boundaries and estimate the average grain size from these curvatures.

In this study, we analyzed the chemical composition of Al-Si cast alloys from microstructural images, using image recognition and machine learning. Binary Al-Si alloys of Si = 1~10 wt% were cast and prepared as reference images in the dataset used for machine learning. The machine learning procedure was constructed with Inception V3 model. Repeated training to relate the microstructural images to their chemical composition was carried out, for up to 10,000 steps, to increase the reliability of the analysis. The peaks of similarity existed in the dataset with chemical compositions corresponding to the known target composition. The heights of the peaks became higher and the distribution of similarity became sharper with further training steps. This means that the weighted average of the chemical composition approached the target composition with increasing training steps. The correctness of the analysis increased with training steps up to 10,000, then was saturated. It was found that the chemical composition outside the dataset range could not be analyzed correctly. Analysis of the compositions between the datasets showed incorrect but reasonable results. The reliability of the chemical composition analysis using machine learning and image recognition developed in this study will increase when a vast range of reference images are collected and verified.

Macrostructural features in continuously-cast stainless steel slabs such as the distribution of equiaxed and columnar grains, and microporosity, are crucial to the post-processes, and the mechanical properties of the final steel products. Among the methods for controlling grain structure during the continuous casting process, the application of electromagnetic stirring (EMS) during casting is known to alter the grain structure of the slab. In this study, we predicted macrostructural features in a continuous-cast stainless steel slab using computer simulation. We approximated the heat transfer system as a one-dimensional heat conduction model with solidification. The effect of EMS was investigated by introducing an effective thermal conductivity of the melt. We assumed that EMS caused turbulent flow, and enhanced the thermal conductivity. Columnar-to-equiaxed transition (CET) and microporosity were estimated with the models proposed by Hunt and Niyama, respectively. The effect of EMS on microporosity was negligible, and microporosity was inevitable in the centerline of the slab. EMS made a second equiaxed region between the outer wall and the center of the slab. Because the effective thermal conductivity was high due to EMS, the temperature field in the slab was distorted, and the Hunt criterion for CET was satisfied in the region. The CET results were compared and verified with cellular automata simulation. Various process parameters including casting speed, superheating, and the position of the EMS module, were examined to predict the macrostructure of the slab.

We attempted to classify the microstructural images of spheroidal graphite cast iron and grey cast iron using a convolutional neural network (CNN) model. The CNN comprised four combinations of convolution and pooling layers followed by two fully-connected layers. Numerous microscopic images of each cast iron were prepared to train and verify the CNN model. After training the network, the accuracy of the model was validated using an additional set of microstructural images which were not included in the training data. The CNN model exhibited an accuracy of approximately 98% for classification of the cast irons. Typically, CNN does not provide bases for image classification to human users. We tried to visualize the images between the network layers, to find out how the CNN identified the microstructures of the cast irons. The microstructural images shrank as they passed the convolutional and pooling layers. During the processes, it seems that the CNN detected morphological characteristics including the edges and contrast of the graphite phases. The mid-layer images still retained their characteristic microstructural features, although the image sizes were shrunk. The final images just before connecting the fully-connected layers seemed to have minimalized the information about the microstructural features to classify the two kinds of cast irons. Matrix phases such as ferrite and pearlite did not show prominent effects on the classification accuracy.