edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 28 NO. 3 18 AUTOMATED END-TO-END AI SOLUTION FOR PHOTON EMISSION MICROSCOPY IN SEMICONDUCTOR FAILURE ANALYSIS Guofeng You, Silambarasan Karuppannan, Kan Sun, Wilson Lee Cheng Hoe, and Grace Tan Qualcomm Global Trading Pte. Ltd., Singapore gyou@qti.qualcomm.com EDFAAO (2026) 3:18-23 1537-0755/$19.00 ©ASM International® INTRODUCTION In semiconductor failure analysis (FA), rapid defect localization is critical for maintaining yield and reliability. Photon emission microscopy (PEM) serves as a standard technique, detecting photons emitted by defects in active circuitry under electrical bias.[1-3] However, continued scaling and increasing routing density make PEM interpretation harder: Emissions are often faint, scenes are congested, and comparison across operating conditions or revisions may require reviewing many frames with subtle differences. As a result, PEM remains a frequent throughput bottleneck in practical FA workflows. In many labs, PEM analysis is still dominated by manual inspection and judgment. This introduces three recurring issues: (1) The process is labor‑intensive and difficult to scale for high‑throughput demand; (2) Interpretation varies across analysts and sessions—especially for weak or ambiguous hotspots; and (3) Correlating image-space findings to CAD/GDS layout context is often disjointed, requiring repeated context switching between imaging tools and layout viewers, which increases cycle time and error risk. Similar automation motivations have been reported for other FA imaging modalities, including lock‑in thermography[4] and inspection/characterization techniques such as SEM and CSAM[5,6]—highlighting the value of repeatable, algorithmic triage in image‑ driven diagnosis. To address these limitations, this article presents an end-to-end AI application for PEM analysis that integrates three capabilities into a single user workflow: automated hotspot detection, quantitative reference‑to‑target comparison, and layout correlation (Fig. 1). The goal is not only to detect hotspots, but to support FA decision‑ making by ranking changes between conditions and translating pixel‑space observations into design context. Building on advances in computer vision and deep learning for microscopy and defect localization, the system is engineered for day‑to‑day use with an intuitive interface and traceable outputs. The remainder of this article describes the detection model, comparison and correlation methodology, and practical results from internal evaluations. METHODOLOGY AND RESULTS AI TRAINING STRATEGY The AI methodology addresses the unique challenges of PEM image analysis through a comprehensive training strategy. The dataset comprises tens of thousands of TIFF images extracted from historical FA databases, captured Fig. 1 End-to-end PEM AI solution workflow showing the integration of hotspot detection, comparison analysis, and layout correlation with optional Avalon system integration.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==