INDUSTRY

Samsung Electronics deploys in-house AI to detect semiconductor defects

by
Lee Jeong-wan
Published : Nov. 17, 2026 - 10:00:49
    • Copy Completed!

View Korean Original

'Defect Agent' now in production use, currently being refined

Combines vision AI, multimodal AI and multimodal reasoning

Cuts defect analysis from hours to minutes

Quality and yield grow more critical as memory chips advance

Company plans to evolve system into a preventive agent

Samsung Electronics' semiconductor (DS) division campus in Hwaseong. [Samsung Electronics]
Samsung Electronics' semiconductor (DS) division campus in Hwaseong. [Samsung Electronics]

Samsung Electronics has developed an in-house AI system called Defect Agent that tracks the causes of microscopic defects arising during semiconductor manufacturing and proposes remedies — cutting analysis work that once took engineers several hours down to a matter of minutes.

According to industry sources Tuesday, Samsung developed Defect Agent earlier this year and is currently refining it. The system was built entirely in-house without relying on any external AI model.

Defect Agent analyzes images of microscopic particles left on wafer surfaces and the defects they cause, then simultaneously identifies the root cause and suggests a fix. Rather than examining wafer images alone, the system integrates manufacturing process data and equipment histories to arrive at a response.

The system combines vision AI for image analysis, multimodal AI for connecting disparate data sources, and multimodal reasoning to infer root causes.

The vision AI component examines a defect's size, shape and surrounding circuit patterns from multiple angles. It uses the microscopic particles that caused the defect to estimate how they originated and to assess their impact on yield.

Building on that analysis, the multimodal AI links the defect image to wafer maps and equipment histories to define the nature of the fault — pinpointing when and where it occurred. The system then forms hypotheses through reasoning, verifies them and narrows down candidate causes to reach a comprehensive interpretation.

Defect Agent integrates analysis and remediation into a single workflow, moving from defect identification through root-cause determination to corrective action. The system can flexibly loop back to earlier steps as circumstances require, boosting overall efficiency.

Since Defect Agent was introduced, tangible changes have emerged on the semiconductor production floor. Defect analysis time has dropped sharply: engineers previously had to examine defect images manually to identify the suspect process step, but the agent now handles that in one pass. Tasks that took several hours are completed within minutes.

The system has also shifted defect response from after-the-fact analysis to early intervention. Where anomalies had previously been detected only after measurement values went out of range, Defect Agent can now catch abnormal signals in data flows early enough to intervene before losses mount.

Reducing such defects has grown more urgent as AI-driven demand has pushed wafer prices sharply higher. Semiconductor demand from AI data centers is rippling through the entire supply chain, driving up prices for related materials and components.

At the same time, as semiconductor processes become ever finer and more sophisticated, even minor defects translate into yield losses. High-performance memory chips such as HBM4E, the seventh-generation HBM, are highly sensitive to minute process variations, making precise quality control across the entire manufacturing process essential.

Samsung plans to evolve Defect Agent into a preventive agent capable of intervening before problems spread. The goal is an autonomous defect-control environment that handles everything from root-cause analysis to solution proposals in a single pass, forming the foundation for yield improvement. The company's strategy is to further sharpen the system's reasoning capabilities so it can respond quickly when similar problems arise.

Samsung is in the midst of a company-wide AI transformation, or AX initiative. The company's policy is to apply AI across all business functions and channel the resources freed up by AX into high-value-added and new-growth areas. Its Device Solutions (DS) division, which oversees the semiconductor business, is no exception.

The DS division is already using an AI coding agent called Code Mate for semiconductor software development.

Usage of Code Mate has been growing steadily, with monthly token consumption reaching approximately 1.45 trillion tokens as of September — 3.5 times the level recorded in May, when use of the coding agent began in earnest.


jeongwan@heraldcorp.com
This content was produced with the assistance of AI translation services.

MOST READ