Recently, a research team helmed by Li Xinxin and Yang Heng from the National Key Laboratory of Sensor Technology at the Shanghai Institute of Microsystem and Information Technology has made remarkable strides in precision manufacturing and intelligent technologies for micro-electromechanical systems (MEMS). The team introduced an innovative artificial intelligence large model framework, grounded in a physically constrained variational level set autoencoder (VLSet-AE). This approach effectively tackles the challenge of swiftly and accurately extracting intricate morphological features during deep reactive ion etching (DRIE) processes.

By integrating physically constrained deep decoding techniques, an adaptive evolutionary recognition mechanism akin to "balloon expansion," and the three-dimensional reconstruction of multi-dimensional spatiotemporal features, the model successfully automates and refines the extraction of nine essential critical dimensions. The average absolute error stands at a mere 3.65%, with an overall recognition accuracy reaching an impressive 94.3%. Furthermore, the VLSet-AE model boasts rapid training capabilities, completing the process in just 20 seconds, and delivers ultra-fast single-image inference at a rate of 1.2 seconds per image, all while maintaining extremely low computational demands (around 4 million parameters and 50MB of memory usage). This performance not only outstrips that of traditional deep learning models but also underscores its immense potential for large-scale industrial implementation. The related research findings have been published in the internationally renowned journal Microsystems & Nanoengineering.