抄録
Top-view in-situ scanning electron microscopy (SEM) images of the friction interfaces between an electron-transparent film and a polyacetal (POM) pin surface revealed the formation of transfer layers, wear debris (including rolling debris), and freestanding layers; however, the relationship between these phenomena and friction forces remains unclear. In this study, friction forces were estimated from SEM images using a convolutional neural network (CNN). The estimated friction forces closely matched the measured values. The specific regions in the SEM images that contributed to CNN friction force estimation were identified using gradient-weighted class activation mapping (Grad-CAM). These regions were primarily located around the contact surface rather than on the contact surface. Furthermore, to investigate the specific contact surface features associated with lower and higher friction forces beyond those observed in the friction test, fake top-view in-situ SEM images were generated with continuous and arbitrary friction forces, including values outside the experimental friction force range, using a continuous conditional generative adversarial network. The generated SEM images successfully reproduced the experimental features corresponding to the friction forces within a range for which sufficient experimental data were available.
