Source-linked AI summary
Feature identification for parameter extraction and defect detection using machine learning
Yan Guo, Helda Pahlavani, Artem Khachaturiants, Khalid Elsayed, Jakob van de Laar, Erik Simons, Niranjan Saikumar, Hamed Sadeghian
TL;DR
High-speed SPM images need automated feature processing for process control, but the data-rich images and range of defects call for dedicated algorithms. The paper presents a synthetic-image and transfer-learning framework using UNet for feature detection; in the contact-hole example, it reports 100% accuracy with no false positives or negatives.
Problem
SPM image processing is less developed for in-line control, while rich 3D data and varied defects require dedicated feature-detection algorithms.
Source: §1 · p. 1Method
A simulator and noise augmenter generate labeled synthetic SPM images for UNet training, which is refined with real tool images through transfer learning.
Source: §1 · pp. 1–2, §3 · p. 3Results
100% accuracy was reported for contact-hole identification, with no false positives or negatives.
Source: §4 · p. 4Takeaways & Limitations
Abstract
from arXiv · showhide
Process control of advanced semiconductor nodes is not only pushing the limits of metrology equipment requirements in terms of resolution and throughput but also in terms of the richness of data to be extracted to enable engineers to finetune the process steps for increased yield. The move towards 3D structures requires extraction of critical dimension parameters from structures which can vary largely from layer to layer. For in-line process control, the necessary automation forces the development of layer and equipment-specific dedicated image processing algorithms. Similarly, with the increase in stochastic defects in the EUV era, detection of defects at the nm scale requires the identification of features captured in low resolution to meet the throughput requirements of HVM fabs, which can again lead to custom algorithm development. With the emergence of ML-based image processing methods, this process of algorithm development for both cases can be accelerated. In this work, we provide the general framework under which the images obtained from high-speed scanning probe microscopy-based systems can be used to train a network for either feature detection for parameter extraction or defect identification.
1. INTRODUCTION
High-speed SPM supports sub-nm 3D imaging, but its automated in-line image processing is less developed than that of other tools. Rich 3D data and shallow, fast scans make feature detection important for both parameter extraction and defect inspection.
1. INTRODUCTION
- SPM image processing must extract relevant information for process control, yet automatic processing is less developed than for other tools.The introduction links this gap to SPM's prior use mainly in R&D rather than in-line control.Source: §1 · p. 1
- Fast scans of high-aspect-ratio structures capture shallow profiles sufficient for defect detection, but not their full depth.The broad range of possible defects also burdens algorithms targeting zero false negatives and minimal false positives.Source: §1 · p. 1
- The paper presents an ML framework for feature detection supporting either metrology parameter extraction or defect identification.The motivation is that SPM's rich 3D data requires dedicated processing for robust critical-dimension extraction.Source: §1 · p. 1
- A synthetic SPM image generator creates a diverse training set with measurement artifacts, followed by refinement using real tool images.This allows preliminary network training without real data, then transfer learning with collected SPM images.Source: §1 · pp. 1–2
2. SYNTHETIC IMAGE GENERATOR
The synthetic image generator creates SPM training images for an EUV contact-hole layer in the ADI condition. A simulator controls structures and defects and generates masks, while a noise augmenter adds SPM artifacts and noise.
2. SYNTHETIC IMAGE GENERATOR
- The example application is an EUV contact-hole layer in the ADI condition, with missing or merging holes identified as inspection targets.Figure 2 illustrates random missing contact holes in synthetic SPM images.Source: §2 · p. 2, Figure 2 · p. 3
- The simulator creates noise-free SPM images with layer-specific structure variation and user-controlled image, feature, and defect parameters.These include pixel size, feature shape, size and pitch, LCDU variation, and the number of randomly introduced defects.Source: §2 · p. 2
- The noise augmenter adds literature-noted SPM and layer artifacts, with user-controlled noise levels to broaden training scenarios.Listed artifacts include streaks or scratches, large particles, and additive white Gaussian noise.Source: §2 · pp. 2–3
- The generator avoids manual labeling by automatically locating features and defects and creating binary segmentation masks.The example images show randomly missing contact holes with the generated masks overlaid.Source: §2 · p. 3, Figure 2 · p. 3
3. UNET FRAMEWORK FOR FEATURE DETECTION
The feature-detection framework uses a UNet encoder-decoder with skip connections for precise feature localization. Training combines equally weighted binary cross-entropy and dice-score losses.
3. UNET FRAMEWORK FOR FEATURE DETECTION
- UNet combines an encoder and decoder, each with convolution operations, and direct copy paths connect corresponding levels for feature localization.The encoder uses ReLU; the decoder upsamples, while the convolution path forms a bottleneck.Source: §3 · p. 3, Figure 3 · p. 4
- Training uses equally weighted binary cross-entropy and dice-score losses.The passage states that the dice loss balances false positives and true negatives.Source: §3 · p. 3
4. RESULTS
The contact-hole UNet achieved 100% accuracy in the reported example, with no false positives or negatives. The section also compares it with thresholding in difficult images and identifies the training parameters used.
- Thresholding performed improperly on very shallow holes and produced false positives on scratch-like SPM artifacts.Figure 5 compares the developed ML method with thresholding in these difficult cases.Source: §4 · p. 4, Figure 5 · p. 5
- Table 1 lists the UNet training parameters used for the contact-hole example.Source: Table 1 · p. 4, §4 · p. 4
- 100% accuracy, with no false positives or negatives, was reported for UNet contact-hole identification.Figure 4 illustrates identification, including missing holes detected by comparing expected hole coordinates with mask coordinates.Source: §4 · p. 4, Figure 4 · p. 5
5. CONCLUSIONS
The conclusion presents a synthetic-image-to-network framework for SPM image processing, using missing contact holes as its example. The authors describe the results as promising for further SPM image-processing developments.
- The framework generates noise-free images with process-related structural variations, adds SPM artifacts and general noise, then uses the images to train an ML network.Source: §5 · p. 5
- Missing contact holes are the example defect used to demonstrate the framework.Source: §5 · p. 5
- The authors call the example results promising and say the framework can be used for several SPM image-processing developments.Source: §5 · p. 5