Adaptive Machine Learning for Semiconductor Defect Screening
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor device manufacturing processes face challenges in effectively screening critical defects due to high nuisance detection rates in optical inspections, especially at advanced technology nodes, where the performance of nuisance filtering techniques is inadequate, leading to significant manual review requirements and risks of missing critical defects until yield is impacted.
Innovation Solution
An adaptive machine learning system is implemented, comprising an adaptive model controller, a defect/nuisance library, and data modeling analytics, which interfaces with SEM review/inspection tools to validate and label defect candidates, generate defect screening models, and update them continuously, using critical signature analytics to improve defect filtering efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If optical inspection is used to detect defects in semiconductor wafers, then defect detection capability is improved, but the number of nuisances (false alarms) increases significantly
Solution Approach 1:
The system implements a feedback loop where inspection results are continuously analyzed, and the nuisance filter is automatically updated based on identified defect patterns. The machine learning model learns from each inspection cycle, adjusting its classification criteria to reduce false alarms while maintaining defect detection accuracy.
Solution Approach 2:
The system dynamically adjusts inspection parameters and filter thresholds based on learned patterns. By changing the classification parameters of the nuisance filter based on machine learning outcomes, the system adapts to different defect types and process variations, reducing nuisance counts while preserving critical defect detection.
2Quantity of substance
If nuisance filtering technique is applied to reduce false alarms, then the number of nuisances decreases, but critical defects may be filtered out as well
Solution Approach 1:
The feedback mechanism ensures that when critical defects are detected or suspected, the system adjusts the nuisance filter parameters to preserve these defect types. The continuous monitoring and adjustment prevent over-filtering by learning from cases where critical defects were initially misclassified.
Solution Approach 2:
The system applies different filtering criteria to different defect types and locations on the wafer. By customizing the nuisance filter parameters for specific defect categories, the system maintains high sensitivity for critical defects while aggressively filtering benign nuisances in other areas.
3Measurement precision
If manual review is performed to identify critical defects, then defect screening accuracy is improved, but productivity decreases due to time consumption
Solution Approach 1:
The machine learning system performs automatic defect classification and nuisance filtering without requiring continuous manual intervention. The system serves itself by automatically updating its models and parameters based on inspection data, eliminating the need for manual review while maintaining high accuracy.
Solution Approach 2:
The system replaces manual review processes with automated machine learning algorithms. By substituting human operators with intelligent software that continuously learns and adapts, the system achieves both high accuracy and maintained productivity without the time constraints of manual inspection.
4Measurement precision
If inspection complexity is increased to detect smaller critical defects, then measurement precision is improved, but the system becomes more vulnerable to process variations
Solution Approach 1:
The system dynamically adapts its inspection parameters and filter thresholds based on real-time process variations. By making the inspection system flexible and responsive to changing conditions, it maintains high precision for small defect detection while automatically adjusting to process variations that would otherwise cause false alarms.
Solution Approach 2:
The machine learning model continuously adjusts inspection parameters based on learned patterns from process variations. By changing parameters dynamically rather than using fixed thresholds, the system maintains sensitivity to small defects while becoming robust to normal process variations through adaptive parameter tuning.
Data Source
AI summary
A system for auto defect screening using adaptive machine learning includes an adaptive model controller, a defect/nuisance library and a module for executing data modeling analytics. The adaptive model controller has a feed-forward path for receiving a plurality of defect candidates in wafer inspection, and a feedback path for receiving defects of interest already screened by one or more existing defect screening models after wafer inspection. The adaptive model controller selects data samples from the received data, interfaces with scanning electron microscope (SEM) review/inspection to acquire corresponding SEM results that validate if each data sample is a real defect or nuisance, and compiles model training and validation data. The module of executing data modeling analytics is adaptively controlled by the adaptive model controller to generate and validate one or more updated defect screening models using the model training and validation data according to a target specification.