TL;DR
This paper introduces a novel framework for detecting lithography hotspots, crucial for improving semiconductor manufacturing yield. By developing a feature space evaluation index and a simplified density-based layout feature, combined with a robust AdaBoost classifier, they achieve high accuracy with near-zero false alarms, outperforming existing methods on the 2012 ICCAD benchmark.
What Problem Does This Paper Address?
As semiconductor devices shrink (low-k1 lithography), the precision required for manufacturing becomes incredibly demanding. Lithography hotspots, which are patterns on the wafer that deviate from the intended design, can lead to defects and reduced manufacturing yield. Detecting these hotspots early in the physical verification phase is critical to minimize costly process optimization and ensure product quality. While lithography simulation offers high accuracy, it's computationally expensive. Existing methods that avoid simulation, such as pattern matching, struggle with detecting unknown hotspot types. Machine learning approaches show promise but face a significant challenge: balancing high detection accuracy with a low false alarm rate. This trade-off is difficult because simple features might lack descriptive power, leading to missed hotspots, while complex features can cause over-fitting and thus, false alarms.
Methods
The proposed framework operates in two phases: a "Learning Phase" and a "Testing Phase."
- Novel Feature Space Evaluation Index: The core innovation lies in a new index designed to quantify the performance of different layout features. This index helps in selecting features that are both simple and effective for classification. It leverages Principal Component Analysis (PCA) to reduce feature dimensionality and then evaluates the linear separability of hotspots from non-hotspots in this reduced space using a Mahalanobis distance.
- Simplified Density-Based Layout Feature: Based on the feature space evaluation index, the authors propose a simplified density-based layout feature. This feature represents layout patterns by the area values within a grid and optimizes its parameters (encoding area size and number of grids) to achieve good linear separability in a low-dimensional space. This contrasts with conventional density-based features that can lead to complex feature spaces.
- Robust AdaBoost Classifier: To handle the simplified yet potentially complex feature space, the paper employs a "Real AdaBoost" classifier. This algorithm uses a series of weak decision tree classifiers, iteratively adjusting sample weights to focus on misclassified instances. The key is that it leverages the probability distribution function of the simplified layout features, allowing it to achieve high accuracy without the over-fitting issues often associated with strong non-linear classifiers like SVMs.
Datasets
The paper evaluates its method using benchmark data from the ICCAD 2012 contest. This benchmark suite includes five test cases (Case 1 to Case 5) with varying technology nodes (32nm and 28nm) and statistics on the number of hotspots (HS) and non-hotspots (NHS). The data is used for both training and testing the classification models.
Key Insights and Ideas
- The "sweet spot" for features: The research highlights that neither overly simple nor overly complex layout features are ideal. The proposed feature space evaluation index helps find a balance, leading to features that are "learning-algorithm-friendly" and enable efficient classification.
- Low-dimensional, high-impact features: By optimizing parameters using the proposed index, the simplified density-based feature can capture essential layout attributes in a low-dimensional space, making subsequent classification more efficient and less prone to over-fitting.
- AdaBoost's strength with simplified features: The paper demonstrates that AdaBoost, when combined with these simplified features, can effectively handle the nuances of layout patterns without succumbing to the false alarm problem that plagues more complex non-linear classifiers. This suggests that for this specific problem, a robust ensemble of weak learners is more effective than a single powerful, complex model.
- The importance of feature space visualization: The 3D visualizations of the feature space (Figure 6) clearly illustrate why the proposed simplified feature is superior. It creates a more distinct separation between hotspots and non-hotspots, making it easier for a classifier to draw a decision boundary.
Key Results
The proposed framework achieved impressive results on the ICCAD 2012 benchmark data:
- High Accuracy: The method consistently achieved accuracy above 95% across most test cases.
- Near-Zero False Alarms: Crucially, the framework reported an average of 0 false alarms for most cases, a significant improvement over previous methods.
- Competitive Runtime: The runtime was comparable to existing Python implementations, with potential for significant acceleration through C++ implementation.
- Outperformance: The paper explicitly states that their detector outperforms other works from the 2012 ICCAD contest in terms of both accuracy and false alarm rate. Table 2 and Table 3 provide detailed comparisons, showing superior performance against methods [18] and [19], and Table 4 demonstrates its advantage over Logistic Regression (LR) and Support Vector Machines (SVM).
Limitations and Future Work
- Parameter Optimization: The paper notes that the parameter 'a' in the feature space index (equation 6) is set to 0.9 based on preliminary experiments. The authors suggest that automatic determination of the optimal value of 'a' based on technology node is a subject for future work.
- Generalization to New Technology Nodes: While the method shows strong performance on the tested benchmark, its adaptability to entirely new technology nodes or significantly different design rules might require further investigation.
- Implementation Efficiency: Although comparable to Python implementations, further optimization through statically typed languages like C++ is mentioned as a path to significant speedup.
Summary updated