Synthetic OES Endpoint Detection for Variable Etch Conditions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional endpoint detection methods in manufacturing processes, such as those used in semiconductor manufacturing, face challenges in accurately determining process endpoints due to variations in substrate designs, process conditions, and chamber efficiencies, often leading to inefficiencies and increased costs from incorrect endpoint determination.
Innovation Solution
Utilizing a trained machine learning model to analyze optical emission spectroscopy (OES) data from a substrate processing operation, generating synthetic OES time trace data of different wavelengths to improve endpoint detection by distinguishing between responsive and non-responsive sensor channels, thereby enhancing endpoint accuracy without the need for additional reference substrates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional endpoint detection methods are used, then process monitoring is performed, but endpoint detection accuracy deteriorates due to variations in substrate designs, process conditions, and chamber efficiencies
Solution Approach 1:
The patent creates a virtual copy of the spectroscopy sensor by training a machine learning model to replicate the sensor's response. The model learns from historical data containing spectroscopy readings, process parameters, and endpoint information, then generates synthetic spectroscopy data that mimics what the actual sensor would detect. This virtual sensor adapts to different substrate designs and process conditions without requiring physical reconfiguration, thereby improving endpoint detection accuracy while maintaining versatility.
Solution Approach 2:
The patent transforms the approach by changing from direct physical measurement to data-driven parameter transformation. The machine learning model processes multiple input parameters including process conditions, substrate characteristics, and historical spectroscopy data to generate transformed spectroscopy readings. This parameter transformation allows the system to adapt to variations in substrate design and process conditions while maintaining consistent endpoint detection accuracy across different scenarios.
2Measurement precision
If multiple sensor channels are used to improve endpoint detection, then measurement accuracy improves, but device complexity increases
Solution Approach 1:
The patent makes the machine learning model universal by training it on diverse historical data from multiple substrate designs, process conditions, and chamber configurations. The single virtual sensor model performs the function of multiple physical sensors by generating synthetic spectroscopy data that accounts for variations across different operating conditions. This multi-functionality allows the system to maintain high endpoint detection accuracy without increasing physical device complexity, as one model replaces what would otherwise require multiple specialized sensors.
3Measurement precision
If reference substrates are used for endpoint detection calibration, then measurement accuracy improves, but manufacturing cost increases
Solution Approach 1:
The patent replaces physical reference substrates with a virtual copy generated by the machine learning model. Instead of consuming actual reference substrates for calibration and endpoint detection, the system uses synthetic spectroscopy data produced by the trained model. This virtual reference approach eliminates the need to manufacture, process, and dispose of physical reference substrates, thereby maintaining high measurement accuracy while significantly reducing material consumption and associated manufacturing costs.
Solution Approach 2:
The patent replaces expensive, consumable reference substrates with a reusable machine learning model. The model, once trained on historical data, can be deployed indefinitely without consuming additional physical materials. This substitution transforms the cost structure from ongoing consumption of expensive reference substrates to a one-time training cost followed by reusable virtual reference generation, thereby reducing manufacturing costs while maintaining endpoint detection accuracy.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach improves endpoint detection accuracy, reduces costs associated with reference substrate processing, and optimizes manufacturing parameters, leading to more efficient and cost-effective production by minimizing defective products and equipment downtime.
Implementation Method 1
a first optical emission spectroscopy (OES) sensor of the plurality of sensors generates first sensor data including a first OES time trace
Data Source
AI summary
A method includes providing, to a trained machine learning model, first OES time trace data from a substrate processing operation. The first OES time trace data is of a first set of wavelengths. The method further includes obtaining, from the trained machine learning model, synthetic OES time trace data of the substrate processing operation, the synthetic OES time trace data being of a second set of wavelengths, different than the first. The method further includes obtaining second OES time trace data from the substrate processing operation of the second set of wavelengths. The method further includes determining, based on the synthetic OES time trace data and the second OES time trace data, a process endpoint for the substrate processing operation. The method further includes performing an action in view of the process endpoint.