Page 44 - Chip Scale Review_May June_2023-digital
P. 44
Achieving automated shmoo results analysis with a
deep learning method
By Chao Zhou [Teradyne Inc.]
T he shmoo plot is a widely- Proposed model introduction To make the tool compatible with
u s e d t e c h n i q u e i n t h e
The structure of this model is a
semiconductor industry for modified AlexNet, proposed by Alex shmoos of different sizes, a 3-level
(4, 2, 1) SPP layer, which is inserted
evaluating product specifications and Krizhevsky [3], which demonstrated between the convolutional layers and
debugging test vectors. It examines that the self-learning features of neural the FC layers, is introduced in this CNN
the characterization of supply voltage networks can surpass manual capabilities. model (Figure 3). The SPP layer is used
and operating frequency, providing The neural network structure of this to convert different shapes of input
valuable information for improving method is shown in Figure 2. A total into a fixed value, which is required as
yield ratios. However, the manual of eight layers of convolutional neural the input size of the subsequent fully-
review process of shmoo plot results, networks (CNNs) are used, including four connected layer. The core principle of
particularly during test with automated layers of convolution (feature extraction), the SPP layer is to use multiple level
t e s t e q u ip m e nt (AT E), i s t i m e - one spatial pyramid pooling (SPP) layer [4] pooling of different sizes (4*4, 2*2 and
consuming and can prolong the test to process the variable size of the input, 1*1 pooling windows are used in this
period, delaying time to market. To two layers of fully-connected (FC) layers method) to process the output of the last
address this challenge, a deep learning (classification) and one output layer. convolutional layer. Then, it combines
based method has been developed and The first four convolutional layers are to the results to obtain three (16*ch, 4*ch
implemented to analyze shmoo results extract the features of the shmoo plot. The and 1*ch) level feature vectors [4].
automatically. This method, developed convolution kernel shape of all layers in Finally, after flattening these feature
using PyTorch, can accurately analyze the network is 3x3, while the shape of the vectors, they are passed to the next fully-
sh moo resu lt s i n a shor ter t i me shmoo plots in the training data set are all connected layer.
compared to manual methods, and 11x11, which means that a large convolution The role of the fully-connected
can be seamlessly integrated into any kernel to extract the pattern is not layer is to do the classification. The
existing test environment. necessary. A small 3x3 window to capture SPP layer is connected to an FC layer
In the field of ATE, there are several a single pixel and its surrounding pixels with 84-channels, at which point the
techniques described in the literature to is sufficient. Padding is used to preserve extracted features output by the SPP
analyze shmoo data, such as decision the pixel information of the shmoo edge to layer pass through two fully-connected
tree, logistic regression, support vector avoid misjudging missing edge pixels. The layers. A one-dimensional array with
machines, and random forest [1-2]. padding size is set to 1, which, coordinated six elements is output as the result, with
These methods are traditional machine with the convolution kernel, can make the these six elements representing the Fail,
learning techniques, which require the height and width of each convolutional Pass, Vol, Freq, Marginal and Hole
results to be manually reviewed, thereby layer’s output equal to 11x11. values, respectively.
limiting their application and resulting in
poor test accuracy in some cases.
Deep learning, deployed in this
new application, is a machine learning
method in which a neural network with
hidden layers is used to simulate the
operation of a human’s neural actions,
extracting the features of the input signal
and executing the classification work
without human intervention. Nowadays,
deep learning applications in computer
vision and natural language processing
far exceed those of traditional machine
learning methods. In this application, a
neural network was adapted to analyze
and interpret the characteristics of a
shmoo plot, significantly improving test
accuracy (Figure 1).
Figure 1: Functional diagram of a deep learning method.
42
42 Chip Scale Review May • June • 2023 [ChipScaleReview.com]