A bayesian framework to integrate knowledge-based and data-driven inference tools for reliable yield diagnoses

This paper studies the issues of designing a Bayesian framework for the reliable diagnosis of various yield-loss factors induced in semiconductor manufacturing. The proposed framework integrates both the results from knowledge-based and data-driven inference tools as input data, where the former resembles expert's knowledge on diagnosing pre-known yield-loss factors while the latter serves for exploring new yield-loss factors. Three modules with specific designs for yield diagnosis applications are addressed: Pre-Process for generating candidate factors and corresponding prior distributions, Bayesian Inference for calculating posterior distributions, and Post-Process for deriving reliable rankings of candidate factors. The final output, a Bubble Diagram with Pareto Frontier, provides visual interpretations on the integral results from data-driven, knowledge-based and Bayesian inference tools. Specific issues addressed in the proposed Bayesian framework provide directions for implementing a real system.