Chongfan Technology

News

All Categories >

Home page

2026

The Shanghai Institute of Optics and Fine Mechanics has made progress in research on deep-learning–enabled optical scattering metrology.

Author:

Recently, the research team led by Researcher Liu Shijie from the High-Power Laser Components Technology and Engineering Department of the Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, in collaboration with Professor Zhu Rihong’s team at Nanjing University of Science and Technology, has made significant progress in the study of deep-learning–enabled optical scatterometry. The relevant findings, titled “Deep learning-enabled optical scatterometry technique for high-precision and non-destructive measurement of grating microstructure parameters,” have been published in the international journal *Optics Express*.

For advanced optical manufacturing, rapid, non‑destructive, high‑precision measurement of grating nanoscale microstructural parameters remains a significant challenge. Traditional physical inversion methods such as RCWA and FDTD entail substantial computational overhead, while microscopic techniques like SEM and AFM suffer from low throughput, high costs, and difficulties in online inspection. Integrating stable optical measurement systems with deep‑learning‑based analysis engines represents a key approach to enhancing the speed and robustness of grating metrology.

The research team developed a dual-beam optical scattering metrology system and proposed an adaptive self-calibrating physics-constrained convolutional neural network, ASPCNN. This system employs a reference optical path to suppress source‑induced fluctuations, while the model extracts multi‑scale spectral features through an adaptive receptive field fusion module and a self‑calibrating residual attention module, and incorporates a physics‑constrained loss function to enhance prediction reliability. Experimental results demonstrate that the method achieves sub‑nanometer measurement accuracy, with an R² exceeding 0.99, and a single‑sample inference time of 9.07 ms, thereby providing robust support for high‑speed, nondestructive characterization of grating microstructural parameters.