Logo image
Back
A scalable multi-modal attention-based framework for Good-Die-In-Bad-Neighborhood screening in semiconductor manufacturing
Journal article   Peer reviewed

A scalable multi-modal attention-based framework for Good-Die-In-Bad-Neighborhood screening in semiconductor manufacturing

Mohammad Ershad Shaik, Abhishek Mishra, Nagarajan Kandasamy and Nur A. Touba
Microprocessors and microsystems, v 124, p105308
Jul 2026
DOI: https://doi.org/10.1016/j.micpro.2026.105308

Abstract

Attention-based neural networks DPPM reduction GDBN Good-Die-In-Bad-Neighborhood Multi-modal learning Semiconductor test quality Test escape detection Wafer-level defect analysis
Outgoing quality optimization in advanced semiconductor manufacturing is increasingly constrained by defects that escape detection under conventional test flows and rule-based screening. Good-Die-in-Bad-Neighborhood (GDBN) techniques exploit spatial defect locality to address this challenge, yet existing learning-based approaches rely predominantly on wafer-level visual features and often exhibit limited robustness under product and process variability. This paper presents a scalable multi-modal attention-based GDBN framework that jointly models localized spatial context and numerical test parametric behavior to enable neighborhood-aware defect screening. The proposed architecture operates on compact 3×3 die-level receptive fields and employs attention-guided feature extraction with late fusion, delivering strong discriminative capability with low computational overhead suitable for high-volume manufacturing. Experiments on the industrial WM-811K dataset show that the proposed multi-modal framework captures up to 2×more test escapes and achieves up to a 22% normalized incremental DPPM reduction relative to state-of-the-art GDBN methods in a single-product setting. Extensive multi-product and leave-one-product-out evaluations further demonstrate that a substantial fraction of these quality gains is preserved across heterogeneous wafer sizes and defect distributions, confirming strong scalability and generalizability. These results establish multi-modal attention-based GDBN as an effective and scalable solution for high-volume, multi-product semiconductor test quality optimization. •Proposes a novel multi-modal attention-based data fusion architecture.•Captures up to 2×more test escapes and a 22% DPPM reduction compared to state-of-the-art GDBN screening methods.•Evaluation across multiple products, demonstrating scalability and yield–quality trade-offs.

Metrics

1 Record Views

Details

Logo image