Classification of Wafer Maps Defect Based on Deep Learning Methods With Small Amount of Data

IEEE 6th International Conference Engineering & Telecommunication – En&T-2019

Introduction: The Purpose of the Research

Improvement of the quality of pattern recognition method in conditions of a deficient amount of labeled experimental data

Work Accomplished

• Method of preparing the composite training dataset:

  • review of typical manufacturing causes of defect patterns;
  • procedure of synthesized wafe maps creation;
  • adaptive configuration of training dataset.

• New learning DCNN strategy:

  • pretrain stage on pure synthetic dataset;
  • main train stage on composite dataset.

• Numerical calculations and results:

  • DCNN model training: VGG-19, ResNet-50, ResNet-34 and MobileNetV2;
  • experimental comparison of models accuracy on different conditions;
  • dependence of classification accuracy on amount of experimental data.

Review of Typical Manufacturing Causes

Source of experimental data

Synthesis of Wafer Maps

Experimental Comparison

Accuracy Specification of the Top DCNN Model

Conclusion

• Proposal of the method of preparing the composite training dataset

• Development of the new learning DCNN model strategy which improve the final result of accuracy by 1% up to 4%

• Experimental accuracy comparison of VGG-19, ResNet-50, ResNet-34 and MobileNetV2 DCNN models for the different ratio of experimental labeled data to synthesized data

• Achievement of 87.8% final classification accuracy with Rₗₛ = 0.05 on the public dataset WM-811K by ResNet-50

• Formative evaluation of needed amount of experimental data to obtain required accuracy