Run-to-Run Recipe Control Using ML Drift Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional run-to-run process control systems in manufacturing systems face challenges in determining which process control variable settings to modify and how much to modify them to optimize substrate processing, leading to variability and defects, which reduces throughput and efficiency.
Innovation Solution
The implementation of a machine learning-based method that identifies data from previous substrate processes, uses a trained model to predict metrology measurement values and detect drift, and updates the process recipe to minimize variability and defects by applying correction factors to process settings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional process control systems modify process recipe settings to optimize substrate characteristics, then substrate characteristics can correspond to target characteristics, but it is difficult to determine which process control variable settings to modify and how much to modify them
Solution Approach 1:
A machine learning model is introduced as an intermediary between process data and control decisions. The model takes process control variable settings and metrology measurement values as input, and outputs predicted drift values that guide process recipe modifications. This intermediary handles the complex analysis work, making the control system more manageable and effective.
Solution Approach 2:
The system implements a feedback loop where metrology measurement values from previous substrate processes are fed back into the machine learning model. The model predicts drift in process characteristics, and this predicted drift information is used to adjust process control variable settings for subsequent substrates, creating a continuous improvement cycle.
2Manufacturing precision
If extensive metrology data collection is performed to optimize process settings, then process optimization accuracy improves, but manufacturing throughput and efficiency decrease due to time consumption
Solution Approach 1:
The machine learning model is trained in advance using historical process data and metrology measurements. Once trained, the model can quickly predict drift values for new substrates without requiring extensive real-time data collection. This preliminary preparation enables fast, accurate process optimization decisions that maintain throughput.
Solution Approach 2:
Instead of collecting and analyzing extensive metrology data for every substrate optimization decision, the system uses the machine learning model to predict drift with limited current process data. The model compensates for the reduced data input by leveraging patterns learned during training, achieving good optimization accuracy with less real-time data collection.
Data Source
AI summary
Data associated with a first process performed for one or more substrates is identified. An amount of drift of a first set of metrology measurement values for the one or more substrates following completion of the first process and/or the second process from target values is determined. Modifications to a recipe for the second process is determined in view of the determined amount of drift. The second process is updated based on the determined one or more modifications.