Bayesian Process Recipe Tuning for Semiconductor Etch Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current manufacturing processes for electronic devices, such as semiconductor production, face inefficiencies in optimizing etch patterning due to the large number of process control variables and the resource-intensive nature of design of experiments (DoE) for tuning process recipes, leading to inconsistent and suboptimal product outcomes.
Innovation Solution
The implementation of a system that uses Bayesian probabilistic models to optimize etch patterning by generating feature models that correlate metrology data with process settings, allowing for the identification of Pareto efficient manufacturing settings and iterative updates to the process recipe, thereby reducing the need for extensive resource-consuming DoE.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional design of experiment (DoE) methods are used to optimize process recipes, then process parameters can be tuned, but time and resources are excessively consumed
Solution Approach 1:
The system performs preliminary actions by training machine learning models on historical process data before actual optimization is needed. The models are pre-trained to understand the complex relationships between process parameters and outcomes, so when optimization is required, predictions can be made rapidly without conducting full DoE experiments each time
Solution Approach 2:
The system creates virtual copies of the physical manufacturing process through digital twins and simulation models. These computational models replicate the behavior of actual equipment, allowing virtual experimentation and optimization without consuming physical resources or time, thereby reducing the need for extensive real-world DoE trials
2Manufacturing precision
If traditional design of experiment (DoE) methods are used to optimize process recipes, then process parameters can be tuned, but resource consumption increases
Solution Approach 1:
The system implements continuous feedback loops where process data is constantly collected, analyzed, and used to update predictions and recommendations. This closed-loop approach allows the system to learn from actual outcomes and progressively improve optimization accuracy without requiring additional resource-intensive experiments, as the feedback mechanism refines existing models rather than demanding new physical trials
Solution Approach 2:
The optimization system becomes self-service by automatically analyzing process data, identifying optimization opportunities, and recommending parameter adjustments without requiring extensive manual intervention or resource allocation for DoE planning and execution. The system serves itself by leveraging its own accumulated knowledge base to drive continuous improvement
3Productivity
If process recipes are tuned with limited substrates for DoE, then some optimization is achieved, but the solution is suboptimal
Solution Approach 1:
The system fundamentally changes the parameters of optimization by transitioning from physical substrate-based experimentation to computational model-based prediction. By changing the fundamental approach from material-intensive DoE to data-intensive machine learning, the system achieves both rapid optimization and high-quality results simultaneously, as model predictions can be generated instantly without physical constraints
Solution Approach 2:
The system introduces machine learning models as intermediaries between process parameters and outcomes. These models act as mediators that translate complex parameter relationships into actionable optimization recommendations, enabling accurate predictions without direct physical experimentation. The intermediary models bridge the gap between limited data and comprehensive optimization
Data Source
AI summary
An electronic device manufacturing system configured to performing, by manufacturing equipment, a first process on a first substrate according to a process recipe, wherein the process recipe comprises a plurality of setting parameters. The system then generates metrology data associated with a plurality of features and inputs the metrology data into one or more Bayesian probabilistic models. The system then receives an output from the one or more Bayesian probabilistic models based on the metrology data and at least one settings parameter of the plurality of setting parameters. The system then updates, based on the output of the one or more Bayesian probabilistic models, the process recipe by modifying at least one setting parameter of the plurality of setting parameters, and performs, by the manufacturing equipment, a second process on a second substrate according to the updated process recipe.