Diagnostic-driven yield engineering under atypical wafer foundry conditions

Abstract

Typically, diagnostic-driven yield engineering consists of two sequential steps- data mining and failure analysis. Data mining seeks early feedback on suspected manufacturing process weakness while failure analysis reveals the physical defect to understand the root cause. However, under atypical conditions such as that in a wafer foundry environment, sufficient information is not available to attain optimal outcomes. Specifically, during volume data analysis, although Root Cause Deconvolution has the ability to predict process weakness using an unsupervised learning algorithm, it only works on random defects. Additionally, GDS layout of IP-secure product might also not be available to a foundry impeding physical failure analysis. Motivated by heightened demands on wafer foundries to deliver faster yield ramp to gain the competitive edge, this paper proposes solutions to overcome these limitations. An enhanced analytical scheme that offers early insights into process weakness (frontend and backend differentiation) regardless of fail mode, and a solution that enables physical failure analysis in the absence of direct access to GDS layout are proposed. An automated approach that captures image snapshots of suspected polygons without compromising confidentiality of the GDS content of IP-protected product is also developed. Experimental results will be presented as an illustration.