Page 49 - Chip Scale Review_May June_2023-digital
P. 49

node,  observation  of  the  output  of   the learning of the middle layer can be
                                           each layer of the model can help us   a great help to artificial intelligence
                                           understand how the neural network   (AI) research. Exploring the black box
                                           learns, which can also help us compose   problem of neural networks is bound to
                                           the neural network structure.      be a future research direction.
                                             From Figures 7-9 it can be found
                                           that the feature maps of the output of   Future work
                                           the convolutional layer, which is close   The new method has been verified on
                                           to the input, can still distinguish the   some test projects where shmoo tests
                                           relationship with the input shmoo. The   currently help engineers save time on
                                           output of the higher layer has higher-  data analysis. One of the future works is
                                           level  features  that  humans  cannot   to optimize the scheme so that the tool
                                           analyze manually, but it is believed that   can adapt to different types of X/Y axes
        Figure 7: Sample shmoo plot.
                                                                              in shmoo plots. The second is to support
                                                                              shmoo center point (test base point)
                                                                              recognition, and the third is to simplify
                                                                              the neural network structure to reduce
                                                                              the number of trainable parameters
                                                                              and overfitting. Lastly, our goal is to
                                                                              expand the training set. Additionally,
                                                                              this method can be used offline or in
                                                                              near real time in conjunction with an
                                                                              ATE system to enable adaptive testing
                                                                              and preemptive troubleshooting based
                                                                              on wafer classification results while the
                                                                              next wafer is being tested.

                                                                              Summary
                                                                                T h i s a r t i c l e f o c u s e s o n t h e
                                                                              development and application of a deep
                                                                              learning  method for automating the
                                                                              analysis of shmoo plots in ATE testing.
                                                                              The article demonstrates that the
                                                                              proposed method can classify the results
                                                                              of shmoo plots accurately, shortening
                                                                              the process from days to minutes, which
                                                                              significantly improves time to market
                                                                              while ensuring high quality levels. The
                                                                              findings suggest that deep learning is a
                                                                              valuable tool for automating the analysis
                                                                              of ATE test data.







        Figure 8: Output of the first CNN layer.






                                                                                CONFERENCE
                                                                                EXHIBITION
                                                                                NETWORKING
                                                                                                       REGISTER!
                                                                                JUNE 26 –28 2023
                                                                                DRESDEN | GERMANY
                                                                                www.semi.org/eu
        Figure 9: Output of the third CNN layer.

                                                                                                             47
                                                               Chip Scale Review   May  •  June  •  2023   [ChipScaleReview.com]  47
   44   45   46   47   48   49   50   51   52