Estimation of dislocation density on GaAs plates using machine vision

The paper presents the results of the application of artificial intelligence (AI) and machine vision technologies for the detection of surface defects. Plates with a crystallographic surface of {100} monocrystalline GaAs grown by the Czochralski method with liquid encapsulation of the melt were analyzed. Based on the YOLOv8 open architecture, a neural network was trained to recognize pits formed as a result of selective etching of single-crystal GaAs plates, and a solution was proposed to automate the calculation of dislocation density based on the identified pits of selective etching. Monochrome images were used for processing by the neural network, the data array at the training stage amounted to about 40,000 objects. It was found that the average density of etching pits (detection objects) is (3 – 7) × 10 4 cm –2 . When trained on a sufficient amount of data, AI and machine vision algorithms are able to recognize target objects with high confidence, including overlapping ones. Continuous counting (over the entire surface of the plate) and further software processing of the results made it possible to obtain a map of the density distribution of dislocation etching pits, as well as lines of density levels (isolines) with reference to its absolute value. It is shown that the use of the continuous counting method using AI and machine vision technologies in comparison with traditional averaging methods for analyzing the structural uniformity of single-crystal GaAs is justified and appropriate. The obtained results can be used for technological control of dislocation density and identification of patterns in the change of the dislocation structure of single-crystal GaAs depending on the growth modes and post-processing of ingots.