For Microelectromechanical Systems (MEMS), etching quality can determine whether a device performs reliably or fails. During DRIE, structures may develop scalloping, bowing, or notching, all of which can distort sidewalls and reduce fidelity. Yet the standard way to evaluate these outcomes remains labor-intensive: engineers must prepare wafers, capture Scanning Electron Microscopy (SEM) images, and manually trace contours, often spending one to two hours on a single image while still facing error rates of 15–20%. Earlier automated methods improved speed but struggled with noisy, low-contrast SEM data and often failed to capture how etched structures evolve with depth. Due to these challenges, there is a need to carry out in-depth research on more accurate and physically meaningful automated analysis of SEM profiles in Deep Reactive Ion Etching (DRIE).
Researchers from the State Key Laboratory of Transducer Technology, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, associated with School of Graduate Study, University of Chinese Academy of Sciences reported (DOI: 10.1038/s41378-025-01105-z) the work in Microsystems & Nanoengineering in 2026. The team developed a Variational Level Set Autoencoder (VLSet-AE) that automates contour recognition in SEM cross-sections of DRIE structures while embedding physical etching constraints into the model itself. The goal was not simply to classify images, but to create a tool that can turn etched profiles into accurate, large-scale process data for more intelligent microfabrication.
What sets VLSet-AE apart is the way it reconstructs etched contours as evolving geometric interfaces rather than simple pixel boundaries. The model was trained on 1,000 SEM images generated from a 16-run orthogonal DRIE experiment and paired with 1,500 etching-rate measurements to learn physically consistent behavior. It extracted nine critical dimensions from etched profiles, including scallop depth, scallop width, scallop radius, profile angle, trench depth, bow width, mid width, and bottom width. The paper reports a low average prediction error of 3.65% and an overall correlation coefficient of 0.998. Errors were as low as 0.56% for profile angle and 2.29% for scallop depth. Compared with seven advanced models, VLSet-AE achieved the highest recognition accuracy at 96%, the shortest training time at 20 seconds, and the fastest inference time at 1.2 seconds.
“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 practical terms, that means the advance is not just a better image-analysis trick. It signals a move toward manufacturing workflows in which inspection data can feed directly into process tuning, simulation, and real-time decision-making with far less human bottleneck.
The broader implications are significant for next-generation MEMS production. A tool that can rapidly and consistently interpret SEM profiles could help manufacturers build larger process databases, connect recipe settings to real structural outcomes, and shorten the path from experimentation to optimization. The authors also note that future work will extend validation to more extreme DRIE conditions, improve robustness under variable image quality, and extract a wider set of process-side and wafer-side parameters. Even at this stage, the study shows how physically informed Artificial Intelligence (AI) can help transform microfabrication from expert-by-expert inspection into scalable, data-rich manufacturing.
###
References
DOI
10.1038/s41378-025-01105-z
Original Source URL
https://doi.org/10.1038/s41378-025-01105-z
Funding Information
This work is supported by the National Key R&D Plan Project (2023YFB3207900).
About Microsystems & Nanoengineering
Microsystems & Nanoengineering is an online-only, open access international journal devoted to publishing original research results and reviews on all aspects of Micro and Nano Electro Mechanical Systems from fundamental to applied research. The journal is published by Springer Nature in partnership with the Aerospace Information Research Institute, Chinese Academy of Sciences, supported by the State Key Laboratory of Transducer Technology.