Machine Learning Outlier Detection for Substrate and Sensor Drift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manufacturing systems face challenges in identifying outlier substrates and defective sensors, leading to incorrect determinations of substrate quality and consistency, which results in unnecessary substrate removal and decreased system efficiency due to inaccurate data from defective sensors.
Innovation Solution
A machine learning model is trained to predict drift in parameter values for substrates, allowing for substrate and sensor health ratings to be determined, thereby identifying outlier substrates and defective sensors without additional metrology measurements, reducing false positives and system latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional sensor-based monitoring is used to determine substrate quality, then the manufacturing process can operate with standard equipment, but defective sensors generate inaccurate data leading to false outlier identification and unnecessary substrate removal
Solution Approach 1:
The patent introduces an intermediary machine learning model that acts as a mediator between the sensor data and substrate quality determination. The model processes sensor readings and predicts substrate quality metrics, filtering out inaccuracies from defective sensors. This intermediary layer prevents false outlier identification while maintaining high reliability in quality assessment.
Solution Approach 2:
The system implements feedback mechanisms where the machine learning model continuously learns from actual substrate outcomes and sensor data correlations. By analyzing patterns between sensor readings and verified quality measurements, the model adapts to identify when sensors are providing inaccurate data, enabling dynamic adjustment and improved reliability over time.
2Measurement precision
If manual inspection methods are used to identify outlier substrates, then equipment complexity remains low, but it becomes difficult for operators to identify outlier substrates and process drifts
Solution Approach 1:
The patent replaces manual operator inspection with an automated machine learning-based detection system. The model analyzes sensor data to identify outlier substrates and process drifts, substituting human judgment with algorithmic analysis that provides higher measurement precision. This automation handles the complexity of multi-parameter analysis that would be difficult for operators to perform manually.
Solution Approach 2:
The system transforms the detection approach by changing from direct manual measurement assessment to predictive parameter analysis. The machine learning model uses sensor readings to predict substrate quality parameters and identify deviations, enabling precise outlier detection through parameter transformation rather than direct observation.
3Measurement precision
If additional metrology measurements are taken to verify substrate quality, then measurement precision increases, but system latency increases and productivity decreases
Solution Approach 1:
The patent applies preliminary action by using the machine learning model to predict substrate quality outcomes before final metrology verification. The model analyzes sensor data during processing to identify substrates that are likely to be outliers, allowing for targeted verification rather than universal measurement. This preliminary assessment reduces the number of time-consuming metrology measurements needed while maintaining high verification accuracy.
Solution Approach 2:
The system uses partial action by performing metrology measurements on only a subset of substrates identified as high-risk outliers by the machine learning model. Rather than measuring all substrates, the model predicts which ones require verification, reducing overall measurement time and system latency while maintaining sufficient quality assurance through selective verification.
Data Source
AI summary
Methods and systems for detecting outliers at a manufacturing system using machine learning are provided. Data collected by a sensors at a manufacturing system during a current process performed for a first set of substrates is provided as input to a trained machine learning model. One or more outputs are obtained from the trained machine learning model. A first amount of drift of a first set of parameter values for the first set of substrates from a target set of parameter values for the first set of substrates is extracted from the one or more outputs. A second amount of drift of each of the first set of parameter values for the first set of substrates from a corresponding parameter value of a second set of parameter values for a second set of substrates processed according to the current process at the manufacturing system prior to the performance of the current process for the first set of substrates is also extracted from the one or more outputs. A substrate health rating is assigned for each of the first set of substrates based on the first amount of drift. A sensor health rating is assigned for each of the sensors at the manufacturing system based on the second amount of drift. An indication of the substrate health rating for each of the first set of substrates and the sensor health rating for each of the sensors are transmitted to a client device connected to the manufacturing system.