Physics-Informed AI for Semiconductor Chamber Recipe Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semiconductor processing recipe generation relies heavily on historical data and engineer intuition, leading to time-consuming and costly experiments, with models lacking contextual relationship to physical properties, resulting in erroneous extrapolations.
Innovation Solution
Implementing a physics-informed machine learning model that integrates domain knowledge and physical system principles, using a reduced order physics model and a regularized objective function to generate chamber attribute data, enabling real-time process space exploration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If physics-informed machine learning models are implemented, then manufacturing precision and reliability are improved, but device complexity increases
Solution Approach 1:
The patent segments the complex physics-informed machine learning model into distinct functional modules: a physics model component that incorporates domain knowledge and physical constraints, a machine learning component that learns from data, and an integration layer that combines them. This modular segmentation allows the system to achieve high manufacturing precision through physics-guided predictions while managing complexity through structured organization of model components.
Solution Approach 2:
The patent transforms the model representation by changing parameters from purely data-driven to physics-informed parameters. This involves incorporating physical constraints, conservation laws, and domain knowledge as regularization terms or structural priors in the loss function, thereby improving manufacturing precision without requiring exponentially more data, while keeping the model complexity manageable through parameter efficiency.
2Manufacturing precision
If extensive physical experimentation is conducted to generate new processes, then manufacturing precision is improved, but loss of time and productivity decrease
Solution Approach 1:
The patent performs preliminary action by pre-training the machine learning model on extensive historical process data and incorporating physics knowledge beforehand. This preliminary training phase allows the model to learn from past experiments and encode physical principles, so that when new process recipes need to be generated, the model can make accurate predictions without requiring extensive new physical experimentation, thereby maintaining manufacturing precision while dramatically improving productivity.
Solution Approach 2:
The patent creates a virtual copy of the physical manufacturing system through the physics-informed machine learning model. This digital twin or virtual model replicates the behavior of the physical system, allowing process exploration and optimization in silico before physical validation. This copying approach reduces the need for repeated physical experimentation while maintaining accuracy, thus resolving the contradiction between manufacturing precision and productivity.
3Measurement precision
If context-specific experimental data is used for model training, then measurement precision is improved within known boundaries, but adaptability decreases when extrapolating beyond boundary conditions
Solution Approach 1:
The patent introduces physics principles as an intermediary between the context-specific training data and the extrapolation task. The physics-informed components act as a mediator that encodes general physical laws and constraints that are valid beyond the training data boundaries. This intermediary guidance allows the model to maintain measurement precision within known boundaries while also adapting to new process spaces through physics-based reasoning, resolving the contradiction between precision and adaptability.
Data Source
AI summary
Embodiments described herein relate to a method that includes implementing a feature extraction process and a feature fusion process from a data set that includes one or more chamber setting data points, where the data set is augmented by a physics attributes model that uses the one or more chamber setting data points to generate chamber attribute data of one or more processing characteristics within a chamber based on physics modeling. In an embodiment, the method further includes implementing a data segmentation process on the data set with a context specific data segmentation module to form a modified data set. In an embodiment, the method may further include training a machine learning model on the modified data set, wherein training the machine learning model includes minimizing a loss function that includes a regularized objective function that includes a term based on physics informed variables.