Machine learning in predicting and explaining failure using class-imbalanced fab data

H. Belyavin, B. Lerner

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

1 Scopus citations

Abstract

Defective lots due to an ion-implantation fab tool producing flash memories can be detected only six weeks after implantation, leading to material waste and financial loss. Machine learning (ML) is used to predict and explain tool failure. For training the ML model, a small and heavily imbalanced database is used with only 8% records of defective lots. It undermines knowledge representation and accurate prediction of defective lots. A methodology is presented that includes correlation analysis, discretization, down sampling, and model optimization using a performance measure, which is imbalance-insensitive. The methodology improves the accuracy of the ML models in classifying defective lots. The models also identify variables related to ion suppression and extraction voltages and currents and thereby may explain the failure.

Original languageEnglish
Title of host publication21st International Conference on Production Research
Subtitle of host publicationInnovation in Product and Production, ICPR 2011 - Conference Proceedings
EditorsTobias Krause, Dieter Spath, Rolf Ilg
PublisherFraunhofer-Verlag
ISBN (Electronic)9783839602935
StatePublished - 1 Jan 2011
Event21st International Conference on Production Research: Innovation in Product and Production, ICPR 2011 - Stuttgart, Germany
Duration: 31 Jul 20114 Aug 2011

Conference

Conference21st International Conference on Production Research: Innovation in Product and Production, ICPR 2011
Country/TerritoryGermany
CityStuttgart
Period31/07/114/08/11

Keywords

  • Bayesian networks
  • Failure prediction
  • Machine learning
  • Neural networks
  • Semiconductor manufacturing

ASJC Scopus subject areas

  • Control and Systems Engineering
  • Computer Science Applications
  • Industrial and Manufacturing Engineering

Fingerprint

Dive into the research topics of 'Machine learning in predicting and explaining failure using class-imbalanced fab data'. Together they form a unique fingerprint.

Cite this