AI-Driven Semiconductor Inspection and Diagnostics Using Design Information

Michael Yu, Thomas Zanon, Rishi Bamb – PDF Solutions 

Introduction 

Advancements in artificial intelligence (AI) are revolutionizing industries, including semiconductor manufacturing. The ability to harness AI for inspection and diagnostics has emerged as a possible game-changing strategy for improving production efficiency, identifying previously undetectable defects, and reducing time-to-market. This blog post explores ways AI, when integrated with semiconductor design information, could contribute to more effective inspection and diagnostic processes. 

This blog post is based on the keynote presentation given by Michael Yu at the December 2024 PDF Solutions AI Executive Conference    

Understanding Semiconductor Inspection Challenges 

Semiconductor manufacturing is inherently complex, involving microscopic structures and intricate designs that demand precision at every stage. One of the most significant challenges lies in inspecting defects that traditional methodologies, such as optical inspection, cannot detect, including subsurface defects, tiny vias, or obscure shorts. 

Traditional inspection methods like raster scanning e-beam techniques are known for their slow throughput, making them insufficient for the high-speed demands of modern production cycles. These challenges call for smarter, faster, and more reliable solutions, that could pave the way for using AI-powered technologies in semiconductor inspection. This blog post explores several areas where we anticipate AI could  help deliver such faster and more reliable testing and inspection. 

Foundation Capabilities Driving AI in Inspection 

Several foundational capabilities would underpin AI-powered inspection processes, each potentially playing a crucial role in optimizing outcomes. These include: 

• eProbe Technology: Unlike traditional raster scanning e-beam methods, eProbe utilizes a point-scan approach, allowing it to jump directly to areas of interest. This innovative scanning method can achieve magnitudes of improvement in throughput. 
• Fire AI and Fuzzy Pattern Algorithms: The Fire AI platform summarizes design geometries and electrical behaviors, enabling more precise defect detection. Its fuzzy pattern algorithm groups layout patterns into fail mode families, improving inspection targeting. 
• Guided Analytics: By incorporating design and scan diagnostic data into guided analytics tools, engineers gain enriched insights into root causes, facilitating faster and more accurate diagnostics. 
  • Automated daily yield summaries with dashboards 
  • Detection of additional root causes previously not possible 
  • Linkage of physical location information from scan diagnosis to layout pattern analysis 
  • Identification of layout sensitivities in yield issues 
• Integration with Siemens EDA Tessent: Siemens’ Tessent platform enhances scan diagnostics, offering insights into failing cells and nets, and enabling fail mode extraction through advanced machine learning algorithms (root cause deconvolution). This partnership exemplifies the synergy between design information and AI-driven analytics. 
  • Scan diagnosis data from Tessent is imported into Exensio 
  • Raw diagnosis reports provide information on failing nets, cells, and physical locations 
  • Machine learning-based “Yield Insight” analyzes die population patterns 

The integrated solution enables more effective failure analysis and diagnosis 

Smart Inspection Recipes Using AI 

AI’s integration into semiconductor inspection will introduce several innovations that enhance both speed and accuracy. Drawing from semiconductor design data, AI algorithms will be able to generate targeted inspection recipes, identify critical areas for defect detection, and optimize diagnostic processes. Below, we outline three key applications where we expect the AI’s impact to be transformational.

1. Random Defectivity Inspection

Random defectivity inspection traditionally involves analyzing a minuscule percentage of critical features due to time constraints. However, with AI, this process could become significantly more efficient. By leveraging design information, AI could optimize scanning locations and guide e-beam inspections toward areas with the highest likelihood of defects.  

The new approach would be: 

• optimizes scan location selection by learning e-beam behavior on product layouts 
• prioritizes inspection of critical wire that are observable by e-beam 
• achieves a much higher rate of critical area inspection within the same time budget 
• scans a significantly larger percentage of observable wire lengths than solutions without AI support 

Selection criteria include prioritizing: 

• Metal lines with good signal-to-noise ratio 
• Lines that are sufficiently long to present meaningful critical area 
• Lines with appropriate grounding characteristics to be observable by e-beam inspection. 

2. Unknown Systematic Defect Detection

Unknown systematic fail modes in semiconductor manufacturing pose a significant challenge, as they often remain undetected without appropriate inspection algorithms. AI could address this challenge by analyzing design patterns and prioritizing potential failure sites based on historical and behavioral data. 

 With the support of AI the approach would be:  

• Perform cartography of the full layout 
• Use fuzzy pattern algorithms to classify patterns into failure mode groups 
• Prioritize inspection budget for unknown systematic patterns 
• With the goal to achieve much higher in systematic failure mode coverage and to capture up to 99% of unknown systematic pattern locations 

3. Scan Test-Driven Inspection

The impact of AI could further be amplified by integrating scan test data with inspection processes. By analyzing scan vectors and identifying failed nets, we anticipate that AI models will be able to create targeted inspection recipes to locate systematic fail modes. This integration will allow engineers to link physical design attributes to observed defects, significantly improving diagnostic accuracy. 

For example, if a specific scan test highlights a failing net, AI should be able to analyze the layout and identify potential failure sites, such as vias or intersecting metal lines. This approach will enable focused investigations, reducing both time and resource expenditure while enhancing overall diagnostics. 

This approach will leverage scan test results to guide inspection: 

• Identifies specific nets associated with scan test failures 
• Analyzes potential systematic failure patterns in and around those nets 
• Optimizes e-beam inspection recipes to target these specific patterns 
• Successfully identifies systematic failures even when their yield impact is 100x less than random defectivity 

Strategic Benefits of AI Implementation 

The integration of AI into semiconductor inspection workflows brings several overarching benefits: 

• Improved Throughput: AI-powered tools like eProbe prioritize critical areas, significantly reducing inspection time while increasing coverage. This enables manufacturers to meet tight production timelines without compromising on defect detection. 
• Lower Costs: By targeting high-priority areas and reducing unnecessary inspections, AI minimizes resource expenditure. Coupled with its ability to enhance diagnostics, this leads to cost-effective operations. 
• Enhanced Accuracy: AI’s ability to analyze design data and behavior patterns results in more precise defect detection and diagnostics, ensuring higher yield rates and lower defects per billion (DPB) metrics. 
• Future-Proofing: With semiconductors powering emerging technologies such as 5G, AI-driven inspection capabilities provide the adaptability and precision necessary to meet evolving demands. 

The Path Forward for Semiconductor Inspection 

The use of AI to integrate design data with inspection and diagnostics could be a paradigm shift. From optimizing random defect inspections to detecting unknown systematic failures and enhancing diagnostics, AI opens the door to the redefinition what is possible in semiconductor manufacturing. 

To harness the full potential of AI, semiconductor manufacturers should consider the following next steps: 

1. Invest in AI Platforms: Platforms like Fire AI and Siemens Tessent offer proven capabilities for integrating design and scan test data with inspection workflows. 
2. Focus on Skill Development: Equip teams with the knowledge and tools required to operate AI-driven systems effectively. 
3. Adopt a Collaborative Approach: Partnerships between design, test, and analytics platforms can unlock synergies that maximize AI’s impact. 

With these strategies, manufacturers can not only solve today’s inspection challenges but also build a competitive edge for the future. 

Conclusion 

By integrating design information with inspection and diagnostics, semiconductor manufacturers will be able to significantly improve defect detection and yield management without waiting for extensive training data. This approach is expected to achieve orders of magnitude improvements in inspection efficiency and systematic defect detection, providing a pathway to implementation of effective AI models in semiconductor manufacturing. 

This integration of design, test, and yield information will represent a significant advancement in semiconductor inspection and diagnostic capabilities, particularly for addressing challenging defects such as undetectable vias and contacts while managing the throughput limitations of traditional inspection methods.