SCIENCE@home

Deep-learning-driven automation of inner and outer object segmentation in SEM/TEM imaging for semiconductor metrology

I. Sanou,J. Baderot,Vincent Barra,Ali Hallal,Léo Mazauric,Johann Foucher

Published 2025 in Journal of Micro/Nanopatterning, Materials, and Metrology

ABSTRACT

Abstract. Background: In the semiconductor industry, accurate and fast segmentation and metrology on scanning electron microscopy (SEM) and transmission electron microscopy (TEM) images are critical for analyzing complex structures. Traditional image analysis often struggles with low contrast, weak or ambiguous boundaries, and the need to distinguish multiple materials within an object. Aim: To address these challenges, we present a deep-learning approach that improves segmentation accuracy and automates metrology for inner and outer objects at advanced manufacturing steps. Approach: We fine-tune the OneFormer architecture with domain-specific adaptations: (i) a combined loss (Boundary, Dice, Binary Cross-Entropy) to sharpen edges and balance classes; (ii) a two-stage training schedule (semantic then panoptic) to stabilize inner/outer delineation; (iii) modality-specific post-processing—shadow handling for SEM and edge enhancement for TEM; and (iv) training on a curated SEM/TEM dataset with expert annotations. Results: On held-out data, our method reaches 89% recall and 80% precision for object detection (baseline from our previous work: 10% recall, 8% precision). For metrology, we obtain dimensionless R2 scores of 0.84 (inner height), 0.92 (outer height), 0.82 (inner width), and 0.90 (outer width), outperforming OneFormer without fine-tuning (R2≤0.47) and prior methods (R2≤0.57). The pipeline requires only a global bounding box initialization, whereas classical snakes need multi-step initialization. Conclusions: These results indicate reliable automated measurement across SEM and TEM, with explicit handling of shadows and weak boundaries. The framework supports semi-automatic operation with user feedback and can be deployed fully automatically after validation, offering a scalable solution for semiconductor metrology.

PUBLICATION RECORD

  • Publication year

    2025

  • Venue

    Journal of Micro/Nanopatterning, Materials, and Metrology

  • Publication date

    2025-10-01

  • Fields of study

    Materials Science, Computer Science, Engineering

  • Identifiers
    DOI 10.1117/1.JMM.24.4.041406
  • External record

    Open on Semantic Scholar

  • Source metadata

    Semantic Scholar

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-51 of 51 references · Page 1 of 1

CITED BY