Deep learning-enabled optical scatterometry technique for high-precision and non-destructive measurement of grating microstructure parameters.

The non-destructive and high-precision measurement of grating microstructure parameters at the nanoscale presents a significant challenge for conventional techniques in advanced optical manufacturing. This paper proposes a novel optical scatterometry system for non-destructive, high-precision measurement of grating microstructure parameters. The system integrates a custom-designed dual-beam scatterometer with a dedicated deep learning architecture, termed ASPCNN (Adaptive Self-calibrating Physics-constrained CNN), for the analysis engine. The instrument employs a reference-beam design for enhanced stability. At the same time, the ASPCNN model enables rapid parameter retrieval from diffraction spectra via an Adaptive Receptive Field Fusion (ARFF) module for multi-scale feature extraction and a Self-Calibrating Residual Attention (SCRA) module for noise-robust representation. A physics-constrained loss function ensures predictions adhere to physical feasibility. Experimental results demonstrate sub-nanometer accuracy, outperforming established models (e.g., ResNet, U-Net) with the coefficient of determination (R2) > 0.99. The analysis engine achieves a remarkable single-sample inference time of 9.07 ms, enabling real-time metrological capabilities. This work establishes a new paradigm for intelligent, high-speed, and non-destructive optical metrology, offering a robust solution for industrial metrology of grating microstructures.