Search for a command to run...
We report results of applying artificial intelligence (AI) and machine vision technologies for detection of surface defects. We analyzed a {100} crystallographic surface of single-crystal GaAs grown by the liquid encapsulated Czochralski method. The YOLOv8 open architecture was used to train a neural network for recognition of pits produced by selective etching of single-crystal GaAs wafers, and we proposed a solution for automating dislocation density counting using observed selective etching pits. For processing by the neural network, we used monochrome images. The data array in the training stage consisted of about 40 000 objects. The average density of etch pits (detection objects) was determined to be (3–7) × 104 cm–2. In the case of training on a sufficient amount of data, AI and machine vision algorithms are capable of recognizing target objects, including overlapping ones, with high reliability. Counting all etch pits (on the entire wafer surface) and subsequent software processing of results allowed us to obtain a dislocation etch pit density distribution map and etch pit density contour lines, with reference to the absolute density value. The use of the complete count method with AI and machine vision technologies, in comparison with conventional averaging methods for analysis of structural uniformity of single-crystal GaAs, has been shown to be justified and reasonable. The results of this study can be used for technological monitoring of dislocation density and gaining insight into general trends in the variation of the dislocation structure of single-crystal GaAs in relation to conditions of ingot growth and post-growth treatment of the ingots.