Semiconductor Chamber ML Model Transfer Using Single-Substrate Metrology
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine-learning models for semiconductor processing equipment are inefficient in adapting to changes in process chambers due to hardware variations and deteriorations, requiring extensive retraining with multiple substrates, which is time-consuming and resource-intensive.
Innovation Solution
The system modifies an existing machine-learning model using transfer learning methods, repurposing it for different or altered process chambers by updating it with new metrology data from a single test substrate, allowing for faster adaptation and reduced resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a machine-learning model is retrained with multiple substrates to adapt to hardware variations and deteriorations, then the model accuracy and reliability are improved, but the time consumption and resource usage increase significantly
Solution Approach 1:
The system performs preliminary actions by collecting and storing process data and metrology data from multiple process chambers during normal operation. This pre-collected data is then used to train the machine-learning model in advance, so when hardware variations or deteriorations occur, the model is already prepared with relevant data patterns, eliminating the need for time-consuming retraining with multiple substrates
Solution Approach 2:
The system creates a virtual copy of the physical process chamber by training a machine-learning model that replicates the chamber's behavior and characteristics. This virtual model can be updated with data from a single test substrate instead of requiring multiple physical substrates for retraining, significantly reducing time and resource consumption while maintaining model accuracy
2Manufacturing precision
If extensive retraining with multiple substrates is performed to adapt to new process chambers, then the manufacturing precision is improved, but the quantity of substrates consumed increases
Solution Approach 1:
The system uses a virtual model that can be updated with data from a single test substrate to achieve the same process optimization that traditionally required multiple substrates. The virtual model captures the essential characteristics and variations of the process chamber, allowing accurate predictions and optimizations without consuming multiple physical substrates
Solution Approach 2:
The system enables self-service by allowing the machine-learning model to adapt to new process chambers using data from a single test substrate that goes through the processing sequence. The model automatically learns from this single substrate's data and adjusts its predictions, eliminating the need for extensive retraining with multiple substrates
3Manufacturing precision
If traditional DoE methods are used to tune and optimize processes, then the process control is improved, but the time required for convergence to desired process increases
Solution Approach 1:
The system replaces the mechanical trial-and-error approach of traditional Design of Experiments with a data-driven machine-learning model. The model uses process data and metrology data to predict optimal process parameters directly, substituting the iterative physical experimentation with computational analysis that converges much faster while maintaining or improving process control
Data Source
AI summary
An electronic device manufacturing system configured to identify, by a processor, a first machine-learning model trained to generate predictive data for a first process chamber and obtain metrology data associated with a substrate produced by a second process chamber. The manufacturing system is further configured to train a second machine-learning model based on the first machine-learning model and the metrology data, wherein the second machine-learning model is trained to generate predictive data associated with the second process chamber.