AI model automates etch profile analysis for faster MEMS manufacturing

Apr 23, 2026

A physics-constrained AI model called VLSet-AE automates feature extraction from DRIE cross-sections with 96 percent accuracy, replacing slow manual SEM analysis in MEMS fabrication.

(Nanowerk News) Researchers at the Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, have developed an artificial intelligence method that automatically analyzes scanning electron microscopy (SEM) images of deep reactive ion etching (DRIE) structures. The model, called Variational Level Set Autoencoder (VLSet-AE), extracts critical geometric features from etched profiles with high precision. VLSet-AE replaces a manual inspection process that typically takes one to two hours per image and carries error rates of 15 to 20 percent. The work was published in Microsystems & Nanoengineering (“AI-driven feature recognition of SEM profiles in deep reactive ion etching based on physics-constrained variational autoencoder”).

Key Findings

In MEMS fabrication, etching quality directly affects device reliability. DRIE, a widely used technique for creating high-aspect-ratio structures, can produce defects such as scalloping, bowing, and notching along sidewalls. These imperfections distort the intended geometry and reduce structural fidelity. Evaluating the resulting profiles has traditionally required engineers to prepare wafer cross-sections, capture SEM images, and manually trace contours, a workflow that is both slow and prone to inconsistency. Earlier attempts to automate this process improved speed but struggled with the noisy, low-contrast conditions typical of SEM data. Many of these methods also treated etched profiles as fixed image patterns, without accounting for how structure geometry shifts at different depths during etching. These limitations left a gap between what automated tools could deliver and what manufacturers needed for reliable quality control. VLSet-AE closes this gap by modeling etched contours as dynamic geometric interfaces shaped by physical removal processes, rather than treating them as flat pixel boundaries to be classified. The model incorporates physical constraints from the etching process into its architecture, enabling it to interpret profile shapes in a way that reflects actual material behavior. This design allows more consistent feature extraction across varying etching conditions. The team trained VLSet-AE on 1,000 SEM images generated from a 16-run orthogonal DRIE experiment, paired with 1,500 etching-rate measurements. From each profile, the model extracted nine critical dimensions: scallop depth, scallop width, scallop radius, profile angle, trench depth, bow width, mid width, and bottom width. The average prediction error across all features was 3.65 percent, with an overall correlation coefficient of 0.998 between predicted and actual values. When benchmarked against seven other advanced models, VLSet-AE delivered the highest recognition accuracy at 96 percent. It also required the least training time at 20 seconds and the shortest inference time at 1.2 seconds, making it suitable for integration into production environments where speed is essential. “This study suggests that SEM analysis may no longer need to remain a slow, manual checkpoint at the end of etching,” the work indicates. “By linking image interpretation to the physics of interface evolution, the method points toward a more dependable way to read complex etched structures at scale.” In practice, VLSet-AE opens a route for inspection results to feed directly into process adjustment, simulation, and real-time quality decisions without requiring extensive human review. Manufacturers could use such a tool to accumulate larger datasets linking recipe parameters to actual structural outcomes, shortening the gap between experimentation and process optimization. The researchers note that future work will extend validation to more extreme DRIE conditions, improve performance under variable image quality, and broaden the range of extractable process and wafer parameters. By embedding physical understanding into automated image analysis, the approach offers a concrete step toward turning MEMS inspection into a scalable, data-driven operation rather than an expert-dependent bottleneck.