edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 28 NO. 3 22 RESULTS Internal evaluations of the deployed system have yielded compelling results. The most significant metric is the reduction in analysis time. A standard manual analysis session, which typically involves scanning images, identifying spots, and manually correlating them with layout, can take upwards of 30 minutes per job. With the AI-driven solution, this time is reduced to a few minutes—a reduction of over 80%. Furthermore, the consistency of analysis has improved markedly. The variability inherent in human interpretation is removed, providing a standardized output that is reproducible across different analysts and shifts. The system has successfully demonstrated high accuracy in detecting hotspots across both raw and overlay streams, and the layout correlation feature has pro- ven invaluable in accelerating root cause identification. CONCLUSION AND DISCUSSION This work demonstrates the successful development and deployment of an AI‑powered ecosystem for photon emission microscopy image analysis. Through the integration of deep learning-based detection models, robust image comparison algorithms, and automated layout correlation, fundamental inefficiencies and key sources of error inherent in traditional failure analysis workflows are addressed. The solution stands as a testament to the practical applicability of AI in semiconductor manufac- turing, moving beyond proof-of-concept to deliver a tangible, high-efficiency tool that empowers FA engineers. The ability to instantly detect, compare, and correlate emission data not only saves valuable time but also enhances the depth and reliability of failure analysis. Internal evalua- tions demonstrate over 80% reduction in analysis time while maintaining consistency and traceability through embedded provenance metadata. As device technologies continue to scale and design complexity increases, automated and intelligent PEM analysis solutions are expected to become increasingly important. Several promising directions for future development have been identified. First, while the current workflow performs object level comparison through hotspot detection and matching, future work will investigate direct image to image comparison models that learn similarity or change representations from paired PEM inputs. This approach may improve sensitivity to subtle differences and reduce reliance on detection thresholds and hand tuned parameters. Second, tighter coupling between AI inference and PEM image acquisition represents a key next step toward guided, semi-autonomous operation of the PEM tool. In this closed loop workflow, hotspot detections and confidence estimates can be used to drive acquisition decisions in real time, including automated region of interest selection, iterative recentering on candidate emission sites, zoom navigation, and objective lens selection to optimize signal capture and spatial resolution. REFERENCES 1. N. Khurana and C.-L. Chiang: “Analysis of Product Hot Electron Problem by Gated Emission Microscopy,” Proc. Int. Relib. Phys. Symp. (IRPS), 1986, p. 189-194. 2. C. Boit, A. Beyreuther, and N. Herfurth: “Photon Emission in Silicon Based Integrated Circuits,” Microelectronics Failure Analysis Desk Reference, 7th Ed., ASM International, 2019, p. 180-195. 3. Z. Song, F. Stellari, and P. Tran: “Photon Emission Intensity Analysis for Leakage Source Identification,” Proc. Int. Symp. Test. Fail. Anal. (ISTFA), 2023, p. 151-154. 4. K.K. Thinn, et al.: “Advancements in Image Pattern Recognition for Lock-in Thermography Hotspot Detection and Classification with Supervised Learning,” Electronic Device Failure Analysis, 25(3), 2023, p. 4-9. 5. E. Dehaerne, B. Dey, and V. Blanco: “Scanning Electron Microscopybased Automatic Defect Inspection for Semiconductor Manufacturing: A Systematic Review,” J. Micro.-Nanopattern., 24(2), 2025, 020901. 6. R. Wilhelmer, F. Laurent, and T. Djuric-Rissner: “AI Image Enhancement Fig. 6 Graphical user interface showing detection, comparison, and image analysis modes with interactive controls for confidence thresholding, IoU filtering, and result visualization.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==