DRAM Image-to-Design Alignment With Robust Offset Clustering

The patent segments the alignment target into multiple instances distributed across the specimen. Instead of relying on a single alignment target, the system identifies and uses multiple instances of alignment targets (e.g., multiple memory cell arrays or logic blocks) to perform alignment. This segmentation reduces the impact of individual target inaccuracies and provides redundant data for more reliable offset calculation.

The patent merges the alignment results from multiple alignment target instances by aggregating their individual offsets. The system calculates offsets for each instance and then combines them through clustering algorithms to determine a final, more accurate align-to-design offset. This merging process leverages the collective information from multiple targets to overcome the limitations of any single target.

The patent performs preliminary actions by first identifying multiple alignment target instances and calculating their individual offsets using multiple alignment methods before final aggregation. The system pre-processes the data by computing offsets for each instance with different methods (e.g., cross-correlation, phase shift), then prepares the data for clustering and outlier rejection. This preliminary computation enables more accurate final offset determination while organizing the complex processing in manageable stages.

The patent implements feedback mechanisms through iterative offset refinement. The system calculates initial offsets, identifies outliers using statistical methods, rejects inaccurate offsets, and then re-calculates the final offset based on the remaining valid data. This feedback loop continues until convergence, ensuring high accuracy while systematically managing the computational complexity through structured iteration.