Use of neural network and genetic algorithm to model scanning electron microscopy for enhanced image of material surfaces

Scanning electron microscope (SEM) is a typical means to take an image of material surfaces. Enhancing the resolution of surface images is complicated by the presence of complex SEM components. SEM characteristics are studied as a function of its component by means of a statistical factor analysis as well as by constructing a neural network prediction model. A face-centered Box Wilson experiment was conducted to collect experimental data. The SEM components examined include an acceleration voltage and a filament current, a working distance, and a magnification. Main effect analysis revealed a much larger impact of the current or the distance than others. A generalized regression neural network (GRNN) was used to build a prediction model of SEM resolution. The model performance was optimized by using a genetic algorithm (GA). An optimized model yielded an improved prediction of 24% over statistical regression model. A higher resolution was achieved by increasing the voltage, the current, and the distance in particular at lower magnification. The SEM resolution was explained by the variation in focal length and the depth of field in view of secondary electrons.