IC Fabrication Time-Series Anomaly Detection With CAE-POD
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semiconductor manufacturing processes face challenges in efficiently monitoring and correcting anomalous signals in complex environments with high product mixes and technology nodes, leading to potential yield and quality issues due to the time-consuming nature of baseline univariate analysis methods, which often miss critical transitional conditions.
Innovation Solution
Implementing a trained convolutional autoencoder (CAE) module coupled with a pointwise outlier detector (POD) to analyze tool trace and response metrics data, enabling efficient anomaly detection and classification in the latent feature space, triggering corrective actions through an out-of-control action plan (OCAP) module.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If baseline univariate analysis methods are used to monitor tool parameters, then the analysis approach is simple to implement, but the detection speed is slow and critical transitional conditions are missed
Solution Approach 1:
The patent replaces traditional mechanical/univariate analysis methods with a neural network-based multivariate analysis system. The CAE-POD architecture processes multiple tool parameters simultaneously, substituting slow sequential univariate analysis with parallel multivariate processing that detects anomalies faster while capturing transitional conditions that univariate methods miss.
Solution Approach 2:
The patent combines multiple analysis components (CAE encoder, latent feature space, POD detector) into a composite analytical system. This composite approach integrates multiple tool parameters and analysis functions into a unified system that achieves both fast detection and comprehensive monitoring, overcoming the limitations of simple univariate methods.
2Productivity
If complex multivariate analysis systems are implemented to detect anomalies faster, then the detection speed and accuracy improve, but the system complexity increases
Solution Approach 1:
The patent segments the complex anomaly detection task into distinct functional modules: the CAE encoder for feature extraction, the latent feature space for compression and representation, and the POD detector for outlier identification. This segmentation allows each component to be optimized independently while working together to achieve fast and accurate detection, managing system complexity through modular design.
Solution Approach 2:
The patent introduces a latent feature space as an intermediary between raw tool parameters and anomaly detection. This intermediate representation compresses and transforms complex multivariate data into a manageable form that the POD detector can efficiently analyze, reducing the computational burden while maintaining detection accuracy.
3Measurement precision
If more tool parameters are monitored simultaneously to improve detection accuracy, then the anomaly detection accuracy improves, but the data processing time increases
Solution Approach 1:
The patent extracts essential features from multiple tool parameters through the CAE encoder, selecting and concentrating the most relevant information in the latent feature space. This extraction process maintains high detection accuracy by focusing on critical patterns while discarding redundant data, thereby reducing processing time despite monitoring multiple parameters.
Solution Approach 2:
The patent transforms raw tool parameters into compressed latent features through the CAE encoding process. This parameter transformation reduces the dimensionality and complexity of the data while preserving the essential patterns needed for anomaly detection, enabling fast processing of multivariate data without sacrificing accuracy.
Data Source
AI summary
An IC manufacturing system including an anomaly detection and classification engine including a trained convolutional autoencoder (CAE) module coupled to a pointwise outlier detector (POD) module that is fitted to a latent feature space associated with the trained CAE module. The trained CAE module is operable to encode a candidate time series signal into a compressed representation in the latent feature space. The POD module is operable to determine the number of compressed datapoints of the compressed representation predicted to be outliers, which may be used in classifying whether the candidate time series signal is an anomalous signal.