Work

Step-Aware Cross-View Transformer for Unsupervised Semiconductor Manufacturing Equipment Anomaly Detection

Year: 2026

Type: article

Abstract: Semiconductor manufacturing requires strict process stability, as subtle equipment degradations often accumulate gradually across consecutive wafers rather than appearing within a single run. Although deep learning approaches have achieved success in modeling process dynamics, most existing methods remain impervious to process steps, operate on wafers independently, and rely on full-scale Transformer architectures that are often too computationally expensive for real-time deployment. Herein, we propose the Step-Aware Cross-View Transformer (SACVT), a framework explicitly aligned with recipe-driven process structure and wafer-to-wafer continuity. First, SACVT introduces a hierarchical step-tokenization mechanism that treats process steps as fundamental modeling units, constraining attention within steps, improving both efficiency and process interpretability. Second, it enforces cross-wafer latent consistency through a dual cross-view decoding objective, enabling early detection of gradual process drift between consecutive wafers. Finally, a Cloud-to-Edge transfer learning strategy allows a general encoder to be pretrained in the cloud while only a lightweight latent-predictor is fine-tuned at the edge, supporting rapid, label-free deployment under industrial IoT constraints. Ablation studies demonstrate that both the learnable Step Summarizer and the Dual Cross-View objective are essential for high-performance anomaly detection, as removing either component results in more than a tenfold increase in the false-alarm rate. These results confirm that step-aware representation learning and bidirectional wafer-consistency enforcement are necessary to suppress spurious alarms while preserving sensitivity to process drift. Evaluated on real-world data, SACVT achieves an F1-score of 0.94 with near-zero false alarms (<0.1%), demonstrating that our structural and practical innovations provide a robust, scalable solution for semiconductor manufacturing fabs.

Source: IEEE Internet of Things Journal

DOI: 10.1109/jiot.2026.3677542

Authors: Sunuwe Kim, Junho Hwang, Sangwook Park

Institutions: Samsung Electronics (South Korea), Seoul National University, Institute of Engineering Research

Language: 영어

Topic: Fault Detection and Control Systems

Subfield: Control and Systems Engineering

Field: Engineering

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: hybrid

Estimated APC paid: $2,645