Contour Recognition and Feature Extraction of Sem Cross-Sectional Profiles in Deep Reactive Ion Etching Based on Physics-Constrained Variational Level Set Autoencoder

This paper reports a physics-constrained variational level set autoencoder (VLSet-AE) for contour recognition and feature extraction in SEM cross-sectional profiles of deep reactive ion etching (DRIE) for the first time. Leveraging scallop layer segmentation along the etching depth, the VLSet-AE dynamically recognizes, and extracts feature critical dimensions, including trench depth, scallop depth, scallop width, and others. By incorporating physical constraints into the decoder, the model ensures that the reconstructed level set function aligns with etching process, improving accuracy and consistency. A three-dimensional feature recognition framework is then established, mapping time, linewidth, and etching depth to enhance temporal and spatial understanding of etching process. Reconstructed scallop segments are systematically assembled into complete profiles, enabling precise quantification of contour features and facilitating 3D morphology simulation. Experimental validation reveals that the proposed model achieves an average error of2.3% in feature recognition, alongside an average accuracy of 94.3% across nine critical dimensions, demonstrating its precision and reliability. The exceptional accuracy in contour recognition and feature measurement significantly advances automated SEM profile analysis, providing a robust tool for optimizing microfabrication processes and enabling scalable, high-precision manufacturing.