Deep Learning Wafer Defect Prediction for Critical Dimension Control

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current defect detection methods in semiconductor manufacturing are inadequate for predicting defect sites and critical dimension measurements, leading to high nuisance rates and inability to determine defect impact on electrical parameters and yield, especially as design rules shrink.

Innovation Solution

A deep learning model is used to predict defect sites and critical dimension variations across wafers by inputting parameters such as defect location, design features, and tool settings, with iterative validation and retraining to achieve detection accuracy thresholds, and generating heat maps for defect and CD variations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional optical or electron beam inspection tools are used to scan hot spots on wafers, then defect detection is performed, but nuisance rates are high and defect impact on electrical parameters cannot be determined

Engineering Contradiction:

Improvedefect detection accuracyVSAvoidnuisance rate

Core Design Contradiction:

ReliabilityVSObject-generated harmful factors

Solution Approach 1:

The patent introduces an intermediary system that correlates defect location data with electrical test data. This intermediary analysis layer processes inspection results and matches them against electrical parameter variations, filtering out nuisance defects that do not correlate with electrical failures while identifying critical defects that do correlate, thereby reducing nuisance rates without sacrificing detection accuracy

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements feedback by using electrical test results to validate and refine defect detection. Electrical measurements from test structures provide feedback about which detected defects are actually critical, allowing the system to learn and improve its ability to distinguish between nuisance and critical defects over time, thereby reducing nuisance rates while maintaining reliability

Inventive Principle:

Principle #23Feedback

2Measurement precision

If inspection sensitivity is increased to detect smaller defects, then more defects are detected, but the population of nuisance defects also increases dramatically

Engineering Contradiction:

Improvedefect detection sensitivityVSAvoidnumber of detected defects

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent extracts only the critical subset of defects from the total detected population by correlating defect locations with electrical failure data. Instead of treating all detected defects equally, the system extracts and focuses attention only on those defects that correlate with electrical parameter variations, effectively filtering out the large population of nuisance defects while maintaining high detection sensitivity

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies local quality by treating different detected defects differently based on their correlation with electrical failures. Rather than uniform treatment of all defects, the system assigns different priorities and levels of concern to different defect locations based on local electrical test results, allowing high sensitivity detection while managing the volume of critical defects separately from nuisance defects

Inventive Principle:

Principle #3Local quality

3Reliability

If process control attempts to eliminate all detected defects, then defect reduction is pursued, but costs and difficulty increase dramatically

Engineering Contradiction:

ImproveyieldVSAvoidprocess control complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the defect population into critical and non-critical categories based on correlation with electrical failures. This segmentation allows process control to focus resources on eliminating only the critical defect subset that actually impacts yield, rather than attempting to eliminate all detected defects. The segmentation reduces process control complexity by prioritizing actions on high-impact defects while accepting the presence of nuisance defects

Inventive Principle:

Principle #1Segmentation

4Measurement precision

If CD measurement data points are taken at particular sites only, then measurement is performed, but data is too infrequent to make meaningful projections

Engineering Contradiction:

ImproveCD measurement capabilityVSAvoidwafer-wide CD variation information

Core Design Contradiction:

Measurement precisionVSLoss of information

Solution Approach 1:

The patent uses machine learning models trained on sparse CD measurement data to generate virtual copies or predictions of CD values at locations where no direct measurements were taken. The model learns patterns from the available measurement points and extrapolates to predict CD variations across the entire wafer, effectively creating a complete wafer map from limited data points and recovering the lost wafer-wide variation information

Inventive Principle:

Principle #26Copying

Data Source

PatentUS11275361B2Systems and methods for predicting defects and critical dimension using deep learning in the semiconductor manufacturing process

Publication Date: 2022.03.15 KLA CORP

AI summary

An initial inspection or critical dimension measurement can be made at various sites on a wafer. The location, design clips, process tool parameters, or other parameters can be used to train a deep learning model. The deep learning model can be validated and these results can be used to retrain the deep learning model. This process can be repeated until the predictions meet a detection accuracy threshold. The deep learning model can be used to predict new probable defect location or critical dimension failure sites.