A through-focus scanning optical microscopy dimensional measurement method based on deep-learning classification...

Area of Science:

• Metrology
• Optical Microscopy
• Machine Learning

Background:

• Through-focus scanning optical microscopy (TSOM) offers economical, non-contact, and nondestructive 3D nanostructure measurement.
• Existing TSOM dimensional measurement methods, like library-matching and machine-learning regression, have limitations in measurement range, environmental sensitivity, subjective feature extraction, and accuracy.

Purpose of the Study:

• To develop an improved TSOM dimensional measurement method using a deep-learning classification model.
• To overcome the limitations of existing TSOM measurement techniques, enhancing accuracy and reducing subjectivity.

Main Methods:

• Implemented ResNet50 and DenseNet121 deep-learning classification models trained on TSOM images.
• Utilized classification results from trained models as dimensional measurement values.
• Evaluated model performance using test images and calculating mean square error (MSE).

Main Results:

• The DenseNet121 model achieved an MSE of 21.05 nm², and the ResNet50 model achieved an MSE of 31.84 nm².
• Both deep-learning classification models demonstrated significantly higher measurement accuracy compared to the machine-learning regression method.
• Accuracy improved notably with an increased number of training linewidths.

Conclusions:

• A deep-learning classification model is a feasible and effective alternative to machine-learning regression for TSOM dimensional measurements.
• The proposed method significantly enhances measurement accuracy for 3D nanostructures.
• This research provides a foundation for future advancements in high-accuracy dimensional metrology.