Evaluation of crystalline structure quality of Czochralski-silicon using near-infrared tomography

Abstract

In this work, three silicon samples are subject to tomographic scans using a 1.6μm laser. The samples were prematurely terminated due to anomalies during the Czhochralski-process. They are taken as analogues of the in situ crystal, where one sample has known aberrant structure in its lowermost 45 mm. The results of the tomographic scans show a distinct difference in transmission profile between the material of known poor mono-crystalline structure and assumed good structure. Three different analysis tools are constructed and applied to quantify the quality of the structure from the results of the tomographic scans. The first two analysis tools are applied as correlation filters constructed from patterns resembling the indicative transmission profiles of high-quality structure, one pattern being an ideal square wave and the other being experimentally determined from the measurements. Both correlation filters yield clear differentiation of low- vs. high-quality material. The final analysis tool is a deep convolutional neural network (deep CNN) evolved from a predetermined architecture configuration using a genetic algorithm. The trained CNN is shown to differentiate the usable high-quality material from the unusable material with a 98.7% accuracy on a testing set of 76 profiles and successfully assigns quality factors to the material that are in good agreement with the correlation filters and previous observations.