Generative adversarial network-based prediction of microhole profile drilled with high-energy electron beam on silicon wafer

Abstract

Drilling with a high-energy electron beam on a semiconductive ceramic substrate is emerging as an effective solution for creating high-aspect-ratio microholes. This method effortlessly surpasses band gaps and facilitates machining with continuous irradiation. However, the inherent brittleness and crystallinity of the semiconductive ceramic substrate hinder handling for quality analysis of the drilled substrate. From experimental drilling, the unique deformation history of the microhole with a high-energy electron beam hints at the potential for predictive modeling for non-destructive analysis of the microhole. In this study, we proposed a conditional generative adversarial network model for the non-destructive analysis of drilled microholes. We collected a limited number of images of hole inlets and cross-sectional holes for training and testing the network model. To effectively build the predictive model with our limited dataset, we introduced a generative adversarial network architecture with embedding a self-attention mechanism with multiple parallel heads. This architecture combines the advantages of the convolutional neural networks and the self-attention mechanism. The proposed architecture showed improvements in training loss and evaluation for image generation compared to the original convolutional neural network. The predictive precision for the inlet diameter, hole straightness, and drilled depth was enhanced by 11.6 %, 8.3 %, and 15 %, respectively. The maximum improvement in predictive accuracy for the inlet diameter, hole straightness, and drilled depth was 36.2 %, 22.98 %, and 58.6 %, respectively. These results indicate that the proposed model not only generated the geometrical profile of the cross-sectional hole but also accurately predicted geometrical dimensions.