Pattern-Aware Intelligence Enables Nondestructive, Rapid Quantification of High-Aspect-Ratio Silicon Etching

High-aspect-ratio (HAR) structures are widely used in microelectromechanical systems (MEMS), electronic devices, and advanced packaging. Accurate characterization of post-etch microstructural morphology is essential for process control and device reliability. However, existing metrology techniques rely on destructive methods and cannot capture the complete three-dimensional profile. This study proposes a pattern-aware intelligent model that transforms pattern-dependent etching behavior into a physical prior, enabling rapid and minimally destructive reconstruction of HAR structures across pattern scales. A YOLO-Pose-based model is developed to automatically extract subpixel geometric features from scanning electron microscopy images, achieving a quantification accuracy of 95.11% with R² >0.98 for key dimensions. By incorporating the pattern-dependent effect in HAR silicon etching as a physical prior, the model creates a topography network that captures etch behavior across feature sizes under identical process conditions. With minimal destructive observations of shallow regions of the wafer, the framework achieves a 93.96% prediction accuracy in the nondestructive retrieval of critical parameters, including etch depth, sidewall angle, and scallop texture, while demonstrating robust cross-scale generalization across varying layout densities. This approach reduces the data acquisition time to 1 min, enabling near-real-time process feedback. It provides a highly interpretable, cost-effective, and intelligent solution for intelligent manufacturing applications.