Semiconductor Process Parameter Setting Using PC-Based EPM Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semiconductor manufacturing processes face challenges in achieving optimal parameter combinations due to incomplete datasets, imbalanced data, and complex correlations between electrical measurement parameters (EPMs), leading to inefficiencies in machine learning and manual sensitivity analysis, and lack of methods to derive optimal EPM combinations.
Innovation Solution
A method and apparatus that derive principal components (PCs) from EPM datasets, perform data component analysis, and use artificial neural networks to balance datasets and determine optimal parameter combinations through data clustering and variance inflation factors, enabling accurate process feedback.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If measurement parameters are sampled to lower processing cost, then processing cost is reduced, but dataset completeness deteriorates resulting in missing values
Solution Approach 1:
The patent creates synthetic copies of process data by generating virtual process conditions and outcomes through AI models. Synthetic process data is generated to supplement sampled measurement data, filling gaps in the dataset without requiring additional physical measurements, thus maintaining data completeness while keeping processing costs low.
2Adaptability or versatility
If various process conditions are split to increase data diversity, then data coverage is improved, but sample balance deteriorates creating imbalanced datasets
Solution Approach 1:
The patent performs preliminary balancing of process condition data before training the AI model. Process condition data is pre-processed to ensure balanced representation across different conditions, and synthetic data is generated in advance to compensate for imbalanced samples, preventing the AI model from being biased toward dominant conditions.
3Manufacturing precision
If manual sensitivity analysis and parameter tuning are performed to achieve optimal EPM combination, then parameter optimization is improved, but development time increases
Solution Approach 1:
The patent replaces manual sensitivity analysis and parameter tuning with an AI-based automated system. The AI model automatically analyzes the relationships between process conditions and EPMs, identifies optimal parameter combinations, and provides tuning recommendations, eliminating the need for time-consuming manual analysis while maintaining or improving optimization accuracy.
4Measurement precision
If statistical correlation between EPMs is considered to find optimal combination, then optimization accuracy is improved, but analysis complexity increases
Solution Approach 1:
The patent introduces an AI model as an intermediary between raw EPM data and optimization results. The AI model automatically handles the complex statistical correlation analysis between multiple EPMs, extracting meaningful relationships and providing optimized parameter combinations without requiring users to manually analyze complex correlations, thus maintaining accuracy while reducing perceived complexity.
Data Source
AI summary
Determining a semiconductor device manufacturing parameter may include determining an EPM (electrical measurement parameters) group that has a correlation in a baseline EPM dataset including EPMs of a device manufactured under a baseline condition, deriving principal components (PCs) corresponding to main correlation axes between EPMs in the EPM group, deriving a PC-based dataset including a baseline PC dataset and a conditional split PC dataset by converting the baseline EPM dataset and a conditional split EPM dataset measured from devices manufactured under conditional splits into a PC domain, determining, using the PC-based dataset, respective PCs which are effectively changed by the conditional splits, obtaining split variation information of the conditional splits, extracting an optimal point capable of optimizing a figure of merit of a semiconductor device within a range of the PC-based dataset, and deriving information for process feedback for realizing the optimal point using the split variation information.