Machine Learning Hotspot Detection in IC Design
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current hotspot detection methods in integrated circuit design face challenges in accurately identifying lithographic hotspots, particularly due to aliasing issues caused by limited sampling size and position, which can lead to misclassification of patterns and reduced detection efficiency.
Innovation Solution
The implementation of a two-scheme machine learning approach using density-based feature encoding with context windows positioned differently, combined with a two-level classification method, where a level-one classifier separates non-hotspots from potential hotspots and a level-two classifier refines the classification to eliminate false positives, addressing aliasing and improving detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If pattern matching-based hotspot detection tools are used, then detection efficiency is improved, but detection coverage is reduced due to inability to detect unseen patterns
Solution Approach 1:
The system performs preliminary sampling of the layout at multiple positions and orientations to collect training data before the actual hotspot detection. This preliminary action enables the machine learning model to learn from diverse pattern examples, improving both detection efficiency and coverage for previously unseen patterns.
Solution Approach 2:
The patent replaces traditional mechanical pattern matching algorithms with a machine learning-based classification system. This substitution allows the system to generalize from training examples rather than relying on pre-defined pattern templates, thereby detecting both known and unseen hotspot patterns effectively.
2Measurement precision
If supervised machine learning techniques are applied to hotspot detection, then detection accuracy for unseen patterns is improved, but computational complexity increases
Solution Approach 1:
The layout is divided into multiple sampling regions at different positions and orientations, with each region processed independently to generate training samples. This segmentation allows the machine learning model to be trained on manageable subsets of data, reducing overall computational complexity while maintaining high detection accuracy.
Solution Approach 2:
The system uses a limited set of strategically selected sampling positions and orientations rather than processing the entire layout exhaustively. This partial action approach provides sufficient training data for accurate hotspot detection while avoiding the prohibitive computational cost of complete layout analysis.
3Speed
If layout pattern truncation is used during sampling, then processing speed is improved, but aliasing occurs causing misclassification
Solution Approach 1:
The patent introduces asymmetric sampling by varying the position and orientation of sampling windows across the layout. This asymmetric approach ensures that truncated patterns still capture sufficient contextual information to distinguish between different layout features, preventing aliasing while maintaining processing efficiency.
Solution Approach 2:
The system adds dimensional diversity by sampling at multiple orientations and positions rather than using a single fixed sampling window. This multi-dimensional sampling approach compensates for the information loss from truncation, allowing accurate classification despite limited sample sizes.
Data Source
AI summary
Aspects of the invention relate to machine-learning-based hotspot detection techniques. These hotspot detection techniques employ machine learning models constructed using two feature encoding schemes. When two-level machine learning methods are also employed, a total four machine learning models are constructed: scheme-one level-one, scheme-one level-two, scheme-two level-one and scheme-two level-two. The four models are applied to test patterns to derive scheme-one hotspot information and scheme-two hotspot information, which are then used to determine final hotspot information.