Assurance Patterns for Mixed-Type Wafer Defect Recognition

Abstract

This internal reference article examines machine-learning support for semiconductor defect inspection through a design-and-assurance lens. It synthesizes the allocated target literature without reporting new experiments, observations, or performance estimates. The analysis treats the practical unit of review as an image source, label system, recognition model, operator action, and production feedback loop. That framing keeps technical mechanisms, evidence quality, user consequences, and institutional controls visible in the same argument. Particular attention is given to how mixed-type predictions can support traceable inspection decisions. The review distinguishes what each cited source directly addresses from the cross-domain principles used for internal comparison. It argues that credible adoption depends on traceable requirements, context-sensitive evaluation, explicit uncertainty, and a documented path for human intervention. The result is a structured reference for teams considering wafer-map triage and manufacturing quality control, especially where label ambiguity, distribution shift, and automation bias could turn a technically plausible component into an unreliable system. The article is intended to support scoping, design review, and evidence planning; it is not a claim of product readiness or an original empirical study.