Feature Contribution Modeling for Lithography Yield Root Cause Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining root causes of yield loss in lithographic patterning processes are inadequate, as they fail to accurately differentiate between various process features' contributions to yield loss due to averaging effects across multiple substrates, leading to incorrect predictions.
Innovation Solution
A two-model approach is employed, using a global neural network model to identify high-quality substrates, followed by a local interpretable model trained on augmented data to determine the specific contributions of process features to yield loss in a selected subset of substrates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a single global correlation model is used to analyze all substrates, then the model can be trained on large amounts of data, but the accuracy of determining feature contributions to yield loss deteriorates due to averaging effects
Solution Approach 1:
The patent divides the substrate population into multiple subsets based on similarity metrics (yield characteristics, process features). Instead of analyzing all substrates uniformly, the system creates segmented groups where each subset is analyzed by its own dedicated correlation model. This segmentation prevents the averaging effect that occurs when dissimilar substrates are mixed, thereby maintaining high accuracy in feature contribution determination while still utilizing substantial training data within each homogeneous subset.
2Measurement precision
If substrates are segmented into subsets for analysis, then the accuracy of root cause identification improves, but the complexity of the analysis system increases
Solution Approach 1:
The patent performs preliminary actions by automatically clustering substrates into subsets based on their yield characteristics and process features before model training. This preliminary segmentation is done using unsupervised learning algorithms that group similar substrates together. By performing this clustering action beforehand, the system prepares the data in an optimal state for subsequent analysis, enabling accurate root cause identification without requiring manual intervention or complex post-processing of mixed substrate data.
Solution Approach 2:
The system dynamically determines the number of subsets and their compositions based on the actual data characteristics. Rather than using a fixed number of subsets, the methodology adjusts the segmentation parameters (number of clusters, similarity thresholds) based on the variance and distribution of yield data and process features. This adaptive parameter adjustment optimizes the balance between model complexity and analytical accuracy for each specific dataset.
Data Source
AI summary
A method of determining a contribution of a process feature to the performance of a process of patterning substrates. The method may include obtaining a first model trained on first process data and first performance data. One or more substrates may be identified based on a quality of prediction of the first model when applied to process data associated with the one or more substrates. A second model may be trained on second process data and second performance data associated with the identified one or more substrates. The second model may be used to determine the contribution of a process feature of the second process data to the second performance data associated with the one or more substrates.