Machine Learning Optical Proximity Correction for Semiconductor Layouts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning technologies face difficulties in securing data for unique patterns of semiconductor devices, leading to challenges in achieving accurate pattern coverage and process reliability during semiconductor manufacturing.
Innovation Solution
A method using machine learning to generate a layout for semiconductor manufacturing by constructing a dataset from unique patterns, performing optical proximity correction (OPC) with a machine learning model, and iteratively refining the layout to reduce computational load and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning is used for layout generation, then computational simplicity is maintained, but pattern coverage accuracy deteriorates due to difficulty in securing data for unique patterns
Solution Approach 1:
The patent performs preliminary actions by generating multiple candidate layouts before final selection, and by pre-processing full chip data to extract representative patterns. This allows the system to build a comprehensive dataset in advance, improving pattern coverage accuracy without requiring additional unique pattern data during the actual machine learning process.
Solution Approach 2:
The patent creates copies of existing layout data by generating multiple candidate layouts with variations. Instead of requiring unique original patterns, the system generates multiple copies and variations of available patterns, using them as training data to improve pattern coverage accuracy for unique patterns that would otherwise be difficult to secure.
2Reliability
If comprehensive data processing is performed to improve pattern coverage, then manufacturing reliability improves, but computational load increases
Solution Approach 1:
The patent segments the comprehensive data processing into distinct stages: full chip data reception, representative pattern extraction, vector generation, dataset creation, and machine learning model generation. This segmentation allows computational tasks to be distributed and managed efficiently, maintaining manufacturing reliability while controlling computational load at each stage.
Solution Approach 2:
The patent extracts only the essential representative patterns from full chip data, and extracts key features to generate vectors. This extraction process removes unnecessary computational overhead while retaining the critical information needed for reliable pattern coverage, thus improving manufacturing reliability without proportionally increasing computational load.
3Productivity
If iterative refinement is performed to reduce computational load, then processing efficiency improves, but convergence time may increase
Solution Approach 1:
The patent performs preliminary actions by pre-generating candidate layouts and pre-extracting representative patterns before the iterative refinement process. This preliminary preparation reduces the computational burden during iteration, improving processing efficiency while minimizing additional convergence time since the foundational work is already complete.
Solution Approach 2:
The patent implements feedback mechanisms where the machine learning model evaluates candidate layouts and provides feedback for iterative refinement. This feedback-driven approach ensures that each iteration meaningfully improves the layout, balancing processing efficiency gains with convergence time by only performing refinements that provide measurable improvements.
Data Source
AI summary
A non-transitory computer-readable medium storing codes that, when executed by a processor, cause the processor to perform operations of receiving full chip data including specific patterns of a first layout, extracting a representative pattern of the first layout from the full chip data, generating a vector of the extracted representative pattern, generating a first data set based on the generated vector, generating a machine learning model by performing machine learning with respect to the first data set, executing an optical proximity correction (OPC) with respect to the specific patterns of the first layout by using the machine learning model, and generating a second layout based on a result of executing the OPC may be provided.