Machine Learning Proximity Correction for Semiconductor Critical Dimensions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Semiconductor manufacturing processes face inaccuracies due to differences between layout data and actual patterns formed, resulting in reduced accuracy and yield, primarily caused by optical proximity effects and variations in critical dimensions.
Innovation Solution
A proximity correction method using machine learning models that generate and process image data from layout data to predict and adjust critical dimensions of patterns, incorporating sensitivity analysis to optimize the correction process, thereby reducing discrepancies between design and actual patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional proximity correction methods are used, then manufacturing process is simpler, but manufacturing precision deteriorates due to inaccuracies in critical dimension prediction
Solution Approach 1:
The patent replaces traditional mechanical/optical proximity correction methods with a machine learning-based prediction system. The machine learning model learns from historical layout data and actual measurement data to predict critical dimensions, substituting complex physical correction processes with intelligent algorithms that analyze patterns and provide correction recommendations.
Solution Approach 2:
The patent introduces measurement data from actual semiconductor substrates as an intermediary between layout design and manufacturing correction. This measurement data serves as feedback that trains the machine learning model, creating a bridge between theoretical design and practical manufacturing outcomes, enabling more accurate predictions.
2Measurement precision
If more sample regions are used for training, then prediction accuracy improves, but processing time increases
Solution Approach 1:
The patent divides the layout data into multiple sample regions, each containing patterns of interest and their surrounding environments. This segmentation allows the machine learning model to process training data in manageable chunks, improving efficiency while maintaining comprehensive coverage of different pattern scenarios for accurate prediction.
Solution Approach 2:
The patent performs preliminary data preparation by generating training datasets from layout data before actual proximity correction is needed. This advance preparation includes creating sample regions, extracting features, and preparing measurement data, so that when correction is required, the machine learning model is already trained and ready to provide rapid predictions.
3Manufacturing precision
If overlapping image data is processed, then comprehensive pattern analysis is achieved, but data processing complexity increases
Solution Approach 1:
The patent extracts relevant features from overlapping image data, such as pattern shapes, sizes, and spatial relationships, rather than processing the entire raw image datasets. This feature extraction reduces data complexity while preserving the essential information needed for accurate critical dimension prediction and pattern analysis.
Data Source
AI summary
A proximity correction method for a semiconductor manufacturing process includes: generating a plurality of pieces of original image data from a plurality of sample regions, with the sample regions selected from layout data used in the semiconductor manufacturing process; removing some pieces of original image data that overlap with each other from the plurality of pieces of original image data, resulting in a plurality of pieces of input image data; inputting the plurality of pieces of input image data to a machine learning model; obtaining a prediction value of critical dimensions of target patterns included in the plurality of pieces of input image data from the machine learning model; measuring a result value for critical dimensions of actual patterns corresponding to the target patterns on a semiconductor substrate on which the semiconductor manufacturing process is performed; and performing learning of the machine learning model using the prediction value and the result value.