ML Model for Semiconductor Pattern Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The semiconductor fabrication process faces challenges in accurately modeling and replicating patterns due to shape transformations during various sub-processes, such as photolithography and etching, which limits the precision of pattern formation and increases the complexity and cost of integrated circuit (IC) manufacturing.
Innovation Solution
A system and method utilizing machine learning (ML) models trained on pairs of design and physical pattern samples to estimate and generate output data defining the shape of patterns, allowing for precise modeling of semiconductor processes and reducing the time and cost of IC design and manufacturing by simulating both design and physical patterns with high accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If machine learning models are used to model semiconductor fabrication processes, then manufacturing precision and productivity are improved, but device complexity increases due to the need for ML infrastructure and trained models
Solution Approach 1:
The patent creates virtual copies of physical fabrication processes through ML models. These models are trained on pairs of design patterns and corresponding physical patterns to learn the transformation relationships, enabling accurate prediction of physical outcomes without repeating expensive physical experiments. This copying approach resolves the contradiction by achieving high precision through virtual modeling while managing complexity through reusable trained models.
Solution Approach 2:
The patent performs preliminary training of ML models using extensive pairs of design and physical patterns before actual fabrication. This preliminary action creates pre-trained models that capture process variations and transformations, allowing rapid and accurate prediction during production. The complexity is paid for once during training, then reused repeatedly for precision predictions.
2Device complexity
If traditional physical modeling methods are used for semiconductor fabrication, then device complexity is reduced, but manufacturing precision deteriorates due to shape transformations in sub-processes
Solution Approach 1:
The patent replaces complex physical/mechanical modeling approaches with machine learning-based computational models. Instead of attempting to physically simulate every sub-process transformation (photolithography, etching, etc.), the system uses ML models trained on actual pattern pairs to predict outcomes. This substitution maintains simplicity while dramatically improving precision by learning from real data rather than physical assumptions.
Solution Approach 2:
The patent transforms the modeling approach by changing from physical parameter-based models to data-driven statistical models. The ML models learn parameter transformations directly from training data, capturing non-linear relationships and process variations that traditional physical models miss. This parameter transformation enables high precision while keeping the system relatively simple through standardized ML architectures.
3Manufacturing precision
If multiple sub-processes are used in semiconductor fabrication to form patterns, then manufacturing precision is improved, but productivity deteriorates due to increased process complexity and time
Solution Approach 1:
The patent performs preliminary training of ML models using extensive pairs of design and physical patterns before actual fabrication. This preliminary action creates pre-trained models that capture process variations and transformations, allowing rapid and accurate prediction during production. The complexity is paid for once during training, then reused repeatedly for precision predictions.
Solution Approach 2:
The patent creates virtual copies of physical fabrication processes through ML models. These models are trained on pairs of design patterns and corresponding physical patterns to learn the transformation relationships, enabling accurate prediction of physical outcomes without repeating expensive physical experiments. This copying approach resolves the contradiction by achieving high precision through virtual modeling while managing complexity through reusable trained models.
Data Source
AI summary
A system for modeling a semiconductor fabrication process includes at least one first processor and at least one second processor. The at least one first processor is configured to provide at least one machine learning (ML) model, which is trained by using a plurality of pairs of images of a design pattern sample and a physical pattern sample. The physical pattern sample is formed from the design pattern sample by using the semiconductor fabrication process. The at least one second processor is configured to provide an input image representing a shape of a design pattern and/or a physical pattern to the at least one first processor and to generate output data defining the physical pattern and/or the design pattern based on an output image received from the at least one first processor.