Optimized Sampling of Angle-Resolved Scatterometry Data Using End-to-End Compressed Learning Model for Nanograss Deficiency Detection

arXiv:2606.07050 · eess.SP · Submitted 2026-06-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Optimized Sampling of Angle-Resolved Scatterometry Data Using End-to-End Compressed Learning Model for Nanograss Deficiency Detection".

Dev: Here's the gist of the paper. Reliable inspection of nanosurfaces is essential to ensure the quality of nanostructure manufacturing The proposed framework integrates a learnable latitude-based sampling layer with a convolutional neural network, allowing sampling and classification to be jointly optimized during training Evaluation results show that the proposed approach achieves high and stable deficiency-level classification performance under different noise conditions Pruning redundant reasoning tokens The paper addresses the data acquisition challenge by proposing an end-to-end compressed learning framework for 5-level vacancy deficiency detection in zinc oxide nanograss using ARS images This framework integrates a learnable latitude-based sampling layer with a convolutional neural network, allowing sampling and classification to be jointly optimized during training The main contributions of this work include proposing an end-to-end CL framework for nanograss deficiency detection using ARS images where classification and sampling strategy optimization are integrated into a single trainable model They also propose a structured latitude-based sampling scheme that exploits the physical characteristics of ARS scattering patterns, which significantly reduces the sampling optimization search space and provides efficient nonuniform sampling using only 180 learnable parameters Furthermore, they develop a trainable sampling layer based on the proposed scheme and optimize it using a straight-through estimator (STE), enabling gradient-based learning of discrete angular sampling masks despite the nondifferentiable sampling operation How the training signal is built The end-to-end CL model consists of two main parts: a layer designed for sampling ARS data and a lightweight CNN to perform deficiency detection The classification part uses a CNN architecture with 16 filters of size 2 × 2, followed by a ReLU activation function and a max-pooling layer with a kernel size of 6 × 6 and stride of 6 The sampling scheme is based on the physical understanding that polar regions show the highest light intensity, leading to non-uniform sampling where more points are placed in regions with higher-frequency patterns To handle the nondifferentiable sampling operation, a straight-through estimator (STE) is employed to approximate the gradient by treating the discrete operation as if it were the identity function Where the gains come from The gains come from leveraging domain-specific information regarding scattering patterns to reduce the search space and optimization parameters in sampling The latitude-based sampling layer assigns different sampling rates for each latitudinal arc, where these rates are learned during training to focus on informative regions The use of a conditional GAN-based ARS data augmentation approach addresses the limited size of the available ARS dataset by generating synthetic samples conditioned on deficiency labels, which resulted in fast convergence within only a few fine-tuning epochs Validation and Performance Analysis The model was evaluated using simulated ARS data from zinc oxide (ZnO) nanosurfaces with different deficiency levels 0%, 10%, 20%, 40%, and 60% Evaluation showed that the classification model is accurate and robust, achieving high accuracy with a very small sampling rate, reaching full-image accuracy using only 5% to 10% of the data The GAN-generated data augmentation was validated using multiple metrics like FID, MMD, KID, and MS-SSIM against ADDA-generated images The learned sampling behavior consistently selects polar regions and distinct latitudinal arcs near the equator as informative areas for classification The performance gap between the CL model and random sampling is substantial, with the CL model reaching 81.21% accuracy with 50 sampling points, substantially outperforming random sampling which only reaches 61.34% The proposed method is efficient, accurate, and suitable for deficiency detection in nanosurface scatterometry The framework successfully demonstrates that the ARS patterns contain a high level of spatial redundancy and only specific regions include most of the class-discriminative information The results confirm that the proposed method is efficient, accurate, and suitable for deficiency detection in nanosurface scatterometry The learned sampling pattern provides interpretability and supports the validity of the sampling design The observed pattern exhibits a high degree of symmetry, demonstrating that the algorithm’s selection strategy is not biased toward one hemisphere or nanosurface orientation The confusion matrices show that most remaining errors are concentrated in the lower deficiency levels, suggesting that mild deficiency levels produce more similar ARS scattering patterns The framework is effective for data augmentation in ARS applications The proposed approach could be extended to nanosurface quality and integrity analysis in nanohole arrays characterized by angle-resolved optical measurements Further work may also consider extending the proposed framework toward industrial inspection pipelines and surface quality assessment under realistic production conditions The paper is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 497286574 The authors are Mehdi Abdollahpour, Carsten Bockelmann, and Armin Dekorsy Page 1 --- Optimized Sampling of Angle-Resolved Scatterometry Data Using End-to-End Compressed Learning Model for Nanograss Deficiency Detection Mehdi Abdollahpour, Student Member, IEEE, Carsten Bockelmann, Member, IEEE, and Armin Dekorsy are with the Department of Communications Engineering, University of Bremen Corresponding author: Mehdi Abdollahpour Page 2 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 2 Page 3 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 3 Page 4 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 4 Page 5 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 5 Page 6 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 6 Page 7 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 7 Page 8 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 8 Page 9 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 9 Page 10 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 10 Page 11 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 11 Page 12 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 12 Page 13 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 13 Page 14 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 14 Page 15 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 15 Page 16 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 16 Page 17 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 17 Page 18 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 18 Page 19 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 19 Page 20 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 20 Page 21 --- ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 21 Page 2606.07050> The paper is titled "Optimized Sampling of Angle-Resolved Scatterometry Data Using End-to-End Compressed Learning Model for Nanograss Deficiency Detection" The single most informative sentence about the paper's subject and finding is Reliable inspection of nanosurfaces is essential to ensure the quality of nanostructure manufacturing The proposed framework integrates a learnable latitude-based sampling layer with a convolutional neural network, allowing sampling and classification to be jointly optimized during training Pruning redundant reasoning tokens Reliable inspection of nanosurfaces is essential to ensure the quality of nanostructure manufacturing The proposed framework integrates a learnable latitude-based sampling layer with a convolutional neural network, allowing sampling and classification to be jointly optimized during training We propose an end-to-end CL framework for nanograss deficiency detection using ARS images where classification and sampling strategy optimization are integrated into a single trainable model The main contributions of this work include proposing an end-to-end CL framework for nanograss deficiency detection using ARS images where classification and sampling strategy optimization are integrated into a single trainable model They also propose a structured latitude-based sampling scheme that exploits the physical characteristics of ARS scattering patterns, which significantly reduces the sampling optimization search space and provides efficient nonuniform sampling using only 180 learnable parameters Furthermore, they develop a trainable sampling layer based on the proposed scheme and optimize it using a straight-through estimator (STE), enabling gradient-based learning of discrete angular sampling masks despite the nondifferentiable sampling operation Validation and Performance Analysis The model was evaluated using simulated ARS data from zinc oxide (ZnO) nanosurfaces with different deficiency levels 0%, 10%, 20%, 40%, and 60% Evaluation showed that the classification model is accurate and robust, achieving high accuracy with a very small sampling rate, reaching full-image accuracy using only 5% to 10% of the data The GAN-generated data augmentation was validated using multiple metrics like FID, MMD, KID, and MS-SSIM against ADDA-generated images The learned sampling behavior consistently selects polar regions and distinct latitudinal arcs near the equator as informative areas for classification The confusion matrices show that most remaining errors are concentrated in the lower deficiency levels, suggesting that mild deficiency levels produce more similar ARS scattering patterns Page 2606.

Title and authors: Rosa: We’re looking at "Optimized Sampling of Angle-Resolved Scatterometry Data Using End-to-End Compressed Learning Model for Nanograss Deficiency Detection" again. The main idea here is using this end-to-end compressed learning framework to handle the data acquisition problem in a way that's much more efficient than traditional methods.

Dev: It’s about linking the sampling and classification directly into one trainable model, which is a big step in how we approach these inspection tasks.

Taro: Essentially, they are proposing an end-to-end compressed learning framework for five-level vacancy deficiency detection in zinc oxide nanograss using ARS images. This means they’ve integrated the sampling and classification strategy optimization into a single trainable model where you optimize both simultaneously during training.

Rosa: That joint optimization is key because it allows the system to learn which parts of the data are most useful for identifying those five deficiency levels, instead of just taking a fixed set of measurements.

Dev: They propose this end-to-end CL framework where classification and sampling strategy optimization are integrated into a single trainable model.

Taro: They also introduce this structured latitude-based sampling scheme that uses the physical characteristics of ARS scattering patterns to significantly reduce the search space for optimization and provides efficient nonuniform sampling using only one hundred eighty learnable parameters.

Rosa: That reduction in parameters is compelling because it makes the process much more manageable, and it’s based on exploiting those physical characteristics of how light scatters in the sphere.

Dev: They are leveraging domain-specific knowledge about scattering patterns to reduce the optimization search space and provide efficient nonuniform sampling with just one hundred eighty learnable parameters.

Taro: Furthermore, they developed this trainable sampling layer based on that scheme and optimized it using a straight-through estimator so they can get gradient learning for discrete angular sampling masks despite the nondifferentiable nature of the operation.

Rosa: That straight-through estimator part is where I think some real engineering muscle is applied to keep the whole training pipeline flowing smoothly.

Dev: It enables gradient-based learning of discrete angular sampling masks even though that sampling operation isn't differentiable, which is a tricky thing to get right in deep learning.

Taro: They used simulated ARS data from zinc oxide nanograss with different deficiency levels, like zero, ten, twenty, forty and sixty percent vacancy deficiencies at the pole of the sphere where the nanosurface is placed at the center of the sphere and light is directed onto it.

Rosa: So they’re testing this framework on controlled simulated data using ADDA images to see how well it performs across all those deficiency levels.

Dev: They used ARS images of zinc oxide nanograss with vacancy deficiencies at deficiency levels of zero, ten, twenty, forty, and sixty percent vacancy deficiencies at the pole of the sphere where the nanosurface is placed at the center of the sphere and light is directed onto it.

Taro: The simulation helps in creating a controlled dataset for evaluating their proposed method for deficiency detection.

The paper's summary: Rosa: To summarize what they actually did, they introduced this end-to-end compressed learning framework specifically to tackle the data acquisition challenge inherent in angle-resolved scatterometry.

Dev: They’re proposing an end-to-end CL framework for nanograss deficiency detection using ARS images where classification and sampling strategy optimization are integrated into a single trainable model.

Taro: The main contributions involve proposing this end-to-end CL framework where classification and sampling strategy optimization are integrated into a single trainable model, which is what they call their main contribution.

Rosa: They also proposed this structured latitude-based sampling scheme that exploits the physical characteristics of ARS scattering patterns to significantly reduce the sampling search space and provides efficient nonuniform sampling using only one hundred eighty learnable parameters.

Dev: They also propose a structured latitude-based sampling scheme that exploits the physical characteristics of ARS scattering patterns to significantly reduce the search space and provide efficient nonuniform sampling using only one hundred eighty learnable parameters.

Taro: They also developed this trainable sampling layer based on that scheme and optimized it using a straight-through estimator so they can get gradient learning for discrete angular sampling masks despite the nondifferentiable operation.

Rosa: And to handle data scarcity, they used a conditional GAN model and its training process to generate synthetic samples conditioned on deficiency labels.

Dev: They also developed this trainable sampling layer based on that scheme and optimized it using a straight-through estimator so they can get gradient learning for discrete angular sampling masks despite the nondifferentiable operation.

Taro: They used simulated ARS data from zinc oxide nanograss with different deficiency levels of zero, ten, twenty, forty, and sixty percent vacancy deficiencies at the pole of the sphere where the nanosurface is placed at the center of the sphere and light is directed onto it.

Rosa: So they’re using this simulation to evaluate their method's ability to detect those five levels accurately in a controlled environment.

Dev: They used ARS images of zinc oxide nanograss with vacancy deficiencies at deficiency levels of zero, ten, twenty, forty, and sixty percent vacancy deficiencies at the pole of the sphere where the nanosurface is placed at the center of the sphere and light is directed onto it.

Taro: This simulation helps in creating a controlled dataset for evaluating their proposed method for deficiency detection.

The paper's improvements: Rosa: When we look at how they improved things, one key improvement is leveraging domain-specific information regarding scattering patterns to reduce the search space and optimization parameters in sampling.

Dev: They gain this by having the latitude-based sampling layer assign different sampling rates for each latitudinal arc that they learn during training to focus on informative regions.

Taro: That learned behavior means the system is focusing its effort on areas with higher-frequency patterns, which captures more essential scattering information than just uniform sampling.

Rosa: They also made the gains from using a conditional GAN based ARS data augmentation approach to address the limited size of the available ARS dataset.

Dev: The use of a conditional GAN based ARS data augmentation approach addresses the limited size of the available ARS dataset by generating synthetic samples conditioned on deficiency labels, which resulted in fast convergence within only a few fine-tuning epochs.

Taro: They also improved robustness under noise conditions, achieving high and stable deficiency-level classification performance under different noise conditions.

Rosa: That stability is important because it means the classification isn't going to wildly fluctuate if there's some background noise in the measurement.

Dev: The model achieves high and stable deficiency-level classification performance under different noise conditions, such as Gaussian noise and speckle noise, demonstrating robustness where the two class scenario remains above ninety-eight percent accuracy.

Taro: They also showed that their learned sampling behavior consistently selects polar regions and distinct latitudinal arcs near the equator as informative areas for classification.

Rosa: That confirms that the physical understanding of ARS scattering patterns leads to a more sensible sampling design, showing where to look most effectively.

Dev: The learned sampling behavior consistently selects polar regions and distinct latitudinal arcs near the equator as informative areas for classification.

Conclusion: Rosa: So wrapping up with the "Optimized Sampling of Angle-Resolved Scatterometry Data Using End-to-End Compressed Learning Model for Nanograss Deficiency Detection," this whole framework is efficient, accurate, and well suited for deficiency detection in nanosurface scatterometry.

Dev: The framework successfully demonstrates that the ARS patterns contain a high level of spatial redundancy and only specific regions include most of the class-discriminative information.

Taro: It shows that mild deficiency levels produce more similar ARS scattering patterns, which explains why most remaining errors are concentrated in the lower deficiency levels.

Rosa: The learned sampling pattern gives us interpretability because it shows exactly which regions are important for classification, and it supports the validity of the sampling design because we can see why we chose those points.

Dev: The observed pattern exhibits a high degree of symmetry, demonstrating that the algorithm’s selection strategy is not biased toward one hemisphere or nanosurface orientation.

Taro: The framework is effective for data augmentation in ARS applications, and they suggest it could be extended to nanosurface quality and integrity analysis in nanohole arrays characterized by angle-resolved optical measurements.

Rosa: It confirms that the ARS patterns have a high degree of symmetry, meaning the algorithm’s selection strategy isn't biased toward one hemisphere or nanosurface orientation.

Dev: The observed pattern exhibits a high degree of symmetry, demonstrating that the algorithm’s selection strategy is not biased toward one hemisphere or nanosurface orientation.

Taro: I think this model is really useful because it offers a way to systematically improve how we acquire data for these types of measurements.

Rosa: It's a solid piece of work showing how to use learning to make the sampling design more informed and less arbitrary, and that’s what we were aiming for.

Dev: Yeah, this framework successfully demonstrates that the ARS patterns contain a high level of spatial redundancy and only specific regions include most of the class-discriminative information.

Taro: We're ready for whatever comes next on arXiv.

Mehdi Abdollahpour, Carsten Bockelmann, Armin Dekorsy

University of Bremen

eess.SP

Submitted: 2026-06-05

Updated: 2026-06-05

Comments: Preprint. 13 pages, 11 figures

Journal ref: IEEE Access, vol. 14, pp. 160680-160696, 2026

DOI: 10.1109/ACCESS.2026.3744028

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 81/100

The gist: Reliable inspection of nanosurfaces is essential to ensure the quality of nanostructure manufacturing The proposed framework integrates a learnable latitude-based sampling layer with a convolutional…

Key concepts

End-to-End CL Framework
This is a single trainable model that handles both data acquisition (sampling) and classification simultaneously. Instead of treating sampling and classification as separate steps, they are trained together to find the best way to select which parts of the ARS image are most important for determining the deficiency level.
Latitude-Based Sampling Layer
This is a learned mechanism that assigns different sampling rates across different latitudes on the ARS sphere. It exploits physical knowledge that regions with higher-frequency scattering patterns, such as polar areas, contain more crucial information about nanograss defects, leading to non-uniform and efficient data selection.
Straight-Through Estimator (STE)
The sampling process involves discrete choices (which points to sample), which is not differentiable for standard neural network training. The STE is a mathematical trick used here to approximate the gradient of this non-differentiable sampling operation, enabling the entire system—including the sampling layer—to be trained using gradient descent.

Terminology

Summary

Reliable inspection of nanosurfaces is essential to ensure the quality of nanostructure manufacturing The proposed framework integrates a learnable latitude-based sampling layer with a convolutional neural network, allowing sampling and classification to be jointly optimized during training Evaluation results show that the proposed approach achieves high and stable deficiency-level classification performance under different noise conditions

Pruning redundant reasoning tokens

The paper addresses the data acquisition challenge by proposing an end-to-end compressed learning framework for 5-level vacancy deficiency detection in zinc oxide nanograss using ARS images This framework integrates a learnable latitude-based sampling layer with a convolutional neural network, allowing sampling and classification to be jointly optimized during training The main contributions of this work include proposing an end-to-end CL framework for nanograss deficiency detection using ARS images where classification and sampling strategy optimization are integrated into a single trainable model They also propose a structured latitude-based sampling scheme that exploits the physical characteristics of ARS scattering patterns, which significantly reduces the sampling optimization search space and provides efficient nonuniform sampling using only 180 learnable parameters Furthermore, they develop a trainable sampling layer based on the proposed scheme and optimize it using a straight-through estimator (STE), enabling gradient-based learning of discrete angular sampling masks despite the nondifferentiable sampling operation

How the training signal is built

The end-to-end CL model consists of two main parts: a layer designed for sampling ARS data and a lightweight CNN to perform deficiency detection The classification part uses a CNN architecture with 16 filters of size 2 × 2, followed by a ReLU activation function and a max-pooling layer with a kernel size of 6 × 6 and stride of 6 The sampling scheme is based on the physical understanding that polar regions show the highest light intensity, leading to non-uniform sampling where more points are placed in regions with higher-frequency patterns To handle the nondifferentiable sampling operation, a straight-through estimator (STE) is employed to approximate the gradient by treating the discrete operation as if it were the identity function

Where the gains come from

The gains come from leveraging domain-specific information regarding scattering patterns to reduce the search space and optimization parameters in sampling The latitude-based sampling layer assigns different sampling rates for each latitudinal arc, where these rates are learned during training to focus on informative regions The use of a conditional GAN-based ARS data augmentation approach addresses the limited size of the available ARS dataset by generating synthetic samples conditioned on deficiency labels, which resulted in fast convergence within only a few fine-tuning epochs

Validation and Performance Analysis

The model was evaluated using simulated ARS data from zinc oxide (ZnO) nanosurfaces with different deficiency levels 0%, 10%, 20%, 40%, and 60% Evaluation showed that the classification model is accurate and robust, achieving high accuracy with a very small sampling rate, reaching full-image accuracy using only 5% to 10% of the data The GAN-generated data augmentation was validated using multiple metrics like FID, MMD, KID, and MS-SSIM against ADDA-generated images The learned sampling behavior consistently selects polar regions and distinct latitudinal arcs near the equator as informative areas for classification The performance gap between the CL model and random sampling is substantial, with the CL model reaching 81.21% accuracy with 50 sampling points, substantially outperforming random sampling which only reaches 61. p. 934%

Conclusion

The proposed method is efficient, accurate, and suitable for deficiency detection in nanosurface scatterometry The framework successfully demonstrates that the ARS patterns contain a high level of spatial redundancy and only specific regions include most of the class-discriminative information The results confirm that the proposed method is efficient, accurate, and suitable for deficiency detection in nanosurface scatterometry The learned sampling pattern provides interpretability and supports the validity of the sampling design The observed pattern exhibits a high degree of symmetry, demonstrating that the algorithm’s selection strategy is not biased toward one hemisphere or nanosurface orientation The confusion matrices show that most remaining errors are concentrated in the lower deficiency levels, suggesting that mild deficiency levels produce more similar ARS scattering patterns The framework is effective for data augmentation in ARS applications The proposed approach could be extended to nanosurface quality and integrity analysis in nanohole arrays characterized by angle-resolved optical measurements Further work may also consider extending the proposed framework toward industrial inspection pipelines and surface quality assessment under realistic production conditions The paper is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 497286574 The authors are Mehdi Abdollahpour, Carsten Bockelmann, and Armin Dekorsy

--- Page 1 ---

Optimized Sampling of Angle-Resolved Scatterometry Data Using End-to-End Compressed Learning Model for Nanograss Deficiency Detection Mehdi Abdollahpour, Student Member, IEEE, Carsten Bockelmann, Member, IEEE, and Armin Dekorsy are with the Department of Communications Engineering, University of Bremen Corresponding author: Mehdi Abdollahpour

--- Page 2 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 2

--- Page 3 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 3

--- Page 4 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 4

--- Page 5 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 5

--- Page 6 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 6

--- Page 7 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 7

--- Page 8 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 8

--- Page 9 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 9

--- Page 10 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 10

--- Page 11 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 11

--- Page 12 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 12

--- Page 13 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 13

--- Page 14 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 14

--- Page 15 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 15

--- Page 16 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 16

--- Page 17 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 17

--- Page 18 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 18

--- Page 19 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 19

--- Page 20 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 20

--- Page 21 ---

ABDOLLAHPOUR et al.: OPTIMIZED SCATTEROMETRY SAMPLING FOR NANOGRASS DEFICIENCY DETECTION 21

--- Page 2606.07050> The paper is titled Optimized Sampling of Angle-Resolved Scatterometry Data Using End-to-End Compressed Learning Model for Nanograss Deficiency Detection

The single most informative sentence about the paper's subject and finding is Reliable inspection of nanosurfaces is essential to ensure the quality of nanostructure manufacturing The proposed framework integrates a learnable latitude-based sampling layer with a convolutional neural network, allowing sampling and classification to be jointly optimized during training

Reliable inspection of nanosurfaces is essential to ensure the quality of nanostructure manufacturing The proposed framework integrates a learnable latitude-based sampling layer with a convolutional neural network, allowing sampling and classification to be jointly optimized during training We propose an end-to-end CL framework for nanograss deficiency detection using ARS images where classification and sampling strategy optimization are integrated into a single trainable model The main contributions of this work include proposing an end-to-end CL framework for nanograss deficiency detection using ARS images where classification and sampling strategy optimization are integrated into a single trainable model They also propose a structured latitude-based sampling scheme that exploits the physical characteristics of ARS scattering patterns, which significantly reduces the sampling optimization search space and provides efficient nonuniform sampling using only 180 learnable parameters Furthermore, they develop a trainable sampling layer based on the proposed scheme and optimize it using a straight-through estimator (STE), enabling gradient-based learning of discrete angular sampling masks despite the nondifferentiable sampling operation

The model was evaluated using simulated ARS data from zinc oxide (ZnO) nanosurfaces with different deficiency levels 0%, 10%, 20%, 40%, and 60% Evaluation showed that the classification model is accurate and robust, achieving high accuracy with a very small sampling rate, reaching full-image accuracy using only 5% to 10% of the data The GAN-generated data augmentation was validated using multiple metrics like FID, MMD, KID, and MS-SSIM against ADDA-generated images The learned sampling behavior consistently selects polar regions and distinct latitudinal arcs near the equator as informative areas for classification The confusion matrices show that most remaining errors are concentrated in the lower deficiency levels, suggesting that mild deficiency levels produce more similar ARS scattering patterns

--- Page 2606.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Optimized Sampling of Angle-Resolved Scatterometry Data Using End-to-End Compressed Learning Model for Nanograss Deficiency Detection: The improved system can perform reliable inspection of nanosurfaces by using a learnable latitude-based sampling layer that exploits physical structures of ARS patterns to reduce the sampling search space and improve convergence.

  2. Joint Optimization of Sampling and Classification: The end-to-end compressed learning framework jointly optimizes the sampling mask to identify informative regions in the scatterometry pattern and the classification network within a single trainable model, making it computationally more efficient.

  3. Fast Convergence via Data Augmentation: The system can achieve fast convergence within only a few fine-tuning epochs when trained on limited data by using a GAN-based augmentation approach that generates synthetic ARS samples conditioned on deficiency labels.

  4. Robust Deficiency Classification Under Noise: The model achieves high and stable deficiency-level classification performance under different noise conditions, such as Gaussian noise and speckle noise, demonstrating robustness where the 2-class scenario remains above 98% accuracy.

  5. Adaptive Sampling Strategy: The learned sampling layer selects physically meaningful regions of the ARS sphere by learning different sampling rates for each latitudinal arc, focusing on regions with higher-frequency patterns to capture essential scattering information more effectively than uniform sampling.

  6. Data Fidelity Validation: Using metrics like FID, MMD, KID, and MS-SSIM, the system can quantitatively verify that GAN-generated ARS images preserve statistical properties and structural characteristics comparable to real ADDA-generated data across various sampling densities.

Sources

Related papers