Skip to content

Defect Agent: Advancing Semiconductor Defect Control through AI

  • mail
Why particle control matters in advanced semiconductor processes

Semiconductor manufacturing involves repeatedly building intricate circuit patterns on a wafer. During this process, minute particles and the defects they cause can directly affect yield. As processes become increasingly miniaturized and highly integrated, defect patterns and their paths of occurrence also become more complex. It is no longer enough to simply ask, "what is visible?" Engineers must also determine "where did it come from?", "which wafers were affected?" and "is the same phenomenon recurring?"

To answer these questions, engineers analyze defect images and wafer map distributions while tracing hundreds of process steps and equipment histories. However, because analysis tools and data are often fragmented, connecting and reviewing these clues require significant time and expertise. In a manufacturing environment where data continues to accumulate, this becomes a bottleneck that slows response.

Defect Agent was developed to address this challenge by connecting images, manufacturing history, and changes over time into a single analytical flow, enabling integration of root-cause tracking with the subsequent response.

 

The reasoning process: A three-layer integrated analysis framework

Defect Agent approaches root-cause analysis through three layers: Vision AI for defect image analysis, Multimodal AI for connecting manufacturing data, and Multimodal Reasoning for root-cause inference.

First, Vision AI goes beyond simple defect classification by analyzing particle size and shape, surrounding circuit patterns, and the area where a particle appears. This reveals what the particle looks like and where it appears, providing a basis for tracing its possible origin and assessing its impact on yield.

However, defect images alone cannot explain why a particle appeared at a particular point in time. Multimodal AI connects this visual information with data such as wafer maps and equipment histories to define the defect phenomenon more specifically. When a particular defect type increases, for example, the system compares its location and timing to narrow down potential suspect equipment. In this way, it clarifies "what happened, where and when”.

Finally, Multimodal Reasoning forms and tests root-cause hypotheses. A defect found on a particular piece of equipment is not immediately assumed to have originated there. Instead, the system compares data from other equipment and examines the timing of status changes. If more evidence is needed, the Agent can use analysis tools, including Tool AI, to review similar cases or changing trends.

By repeating Observation → Hypothesis → Additional Analysis → Re-evaluation, Defect Agent connects different types of information into a single reasoning flow, helping interpret causes and context that may be difficult to identify from individual metrics alone.

 

Integrating analysis and action: End-to-end workflow

Defect control does not end with root-cause determination. Detection, analysis, response planning, action, and result verification must work as a connected process. If data must be retrieved and reinterpreted at every stage, the overall response remains slow regardless of how quickly each individual AI performs.

Defect Agent divides this process into six Task Agents—Phenomenon Analysis, Anomaly Detection, Suspect Analysis, Root Cause Determination, Countermeasure Establishment, and Action Control—and connects them through an Agentic Workflow. Rather than following a rigid pipeline, the Agents can select the tools needed for a given situation or return to an earlier stage when additional information requires reassessment. This creates a flexible End-to-End Workflow that connects analysis, judgment, response planning and action.

 

Defect Agent
A Multimodal Reasoning Agent turns particle images, wafer maps and manufacturing history
into verified root causes and actions

A three-panel infographic shows multimodal evidence flowing through an AI agent workflow, from image and data inputs to evidence-backed decisions, with charts and diagrams.
A three-panel infographic shows multimodal evidence flowing through an AI agent workflow, from image and data inputs to evidence-backed decisions, with charts and diagrams.

 

How Defect Agent changes defect response

Defect Agent changes defect response in three key ways.

First, analysis time can be significantly reduced. Previously, tasks required engineers to manually select wafer maps and defect images and use multiple AI tools to identify suspect processes, which could take hours. Defect Agent can perform these tasks within minutes, allowing engineers to focus on validating suspect equipment candidates and making final judgments.

Second, the process is shifting from post-event analysis toward earlier response. Instead of recognizing a problem only after a measurement exceeds a threshold, the system analyzes changes in data to identify abnormal signals earlier and support intervention before losses expand.

Third, response capability is becoming less dependent on individual experience and more system-driven. Six specialized Task Agents cross-check anomalies, change histories, and similar cases, helping less-experienced engineers perform more consistent End-to-End defect responses.

 

From reactive response to preventive control

In semiconductor manufacturing, production continues even while the root cause of a defect is being investigated. Reducing the time from anomaly detection to root-cause judgment and corrective action is therefore critical. By connecting multiple clues across defect images and manufacturing data, Defect Agent supports an Autonomous Defect Control environment that extends from root-cause analysis to solution recommendation. This can provide a foundation for yield improvement while helping reduce gaps in response capability caused by differences in engineering experiences.

The next step is to turn accumulated reasoning and response experience into reusable manufacturing knowledge, allowing proven resolution paths to be applied more quickly when similar problems arise. Ultimately, the goal is to move beyond post-event root-cause tracking toward Preventive Defect Control, where early signals are identified before issues spread.

What begins with the analysis of an individual particle can evolve into a system that connects judgment and response across the defect-control process and improves through accumulated manufacturing knowledge. This is the future of semiconductor manufacturing AI that Defect Agent is working toward.

 


* The contents of this page are provided for informational purposes only. No representation or warranty (whether express or implied) is made by Samsung or any of its officers, advisers, agents, or employees as to the accuracy, reasonableness or completeness of the information, statements, opinions, or matters contained in this page, and they are provided on an "AS-IS" basis. Samsung will not be responsible for any damages arising out of the use of, or otherwise relating to, the contents of this page. Nothing in this page grants you any license or rights in or to information, materials, or contents provided in this document, or any other intellectual property.
 
* The contents of this page may also include forward-looking statements. Forward-looking statements are not guarantees of future performance and that the actual developments of Samsung, the market, or the industry in which Samsung operates may differ materially from those made or suggested by the forward-looking statements contained in this page.
 
* All product specifications reflect internal test results and are subject to variations by user's system configurations. Actual performance may vary depending on use conditions and environment.
 
* All images shown are provided for illustrative purposes only and may not be an exact representation of the product. All images are digitally edited, modified, or enhanced.