Page 33 - ChipScale_Sep-Oct_2020-digital

A deep-learning solution for heterogeneous

package inspection

By Shahab Chitchian [INTEKPLUS Corporation Ltd.]
H eterogeneous integration and bump damage detection capabilities (objects) from background. It is considered

through the use of system
calculate the probability value of a target
in package (SiP) technology are presented. By mitigation of under- a pixel-level classification algorithm to
and over-rejection for critical defect
has been effectively adopted by the mobile modes, our AI solution results in yield feature for all pixels in the image, then
industry. Recently, chiplet packaging has improvement for our semiconductor to classify it as the target object if the
become a key technology to continue customers, which means significant cost probability is greater than the threshold
Moore’s law by improving yield and reduction for such large form factor and value set by the user. A single model can
reducing total package/product cost. expensive multi-chip packages. be trained to segment several objects, but
In this article, the latest heterogeneous the performance is usually poor. So, our
inspection results covering our deep- The deep-learning process segmentation model classifies pixels into
learning inspection process are reviewed. To understand the AI inspection binary classes of defects and background.
First, we introduce our deep-learning process and results in detail, our AI On the other hand, the classification
algorithms for two steps of segmentation process is briefly summarized below in determines different features (objects) in
and classification, followed by their five sections. The sections are: 1) deep- the image. It is used to classify features
training methods. In the second part learning algorithms; 2) model training with the highest probability related to each
of the article, the integration of deep and validation; 3) deep-learning combined class. Therefore, it has higher performance
learning and machine vision is presented. algorithms and results; 4) second- for multiple classes in one model compared
Furthermore, we show that by applying inspection process and results; and 5) to segmentation. Classification is applied
segmentation and classification in deep-learning deployed on edge computing when it is necessary to distinguish among
series, different defect modes can be vs. cloud computing. different features, e.g., in the case of bump
distinguished and within each mode, Deep-learning algorithms. Our deep- damage to determine a defective bump
different classes of true defects and learning approach consists of segmentation from other non-defective (overkill) bumps.
overkills can be differentiated. Lastly, and classification. On the one hand, Figure 1a shows our segmentation
some case studies including micro-crack segmentation distinguishes target features st e ps a nd exa mple defe ct s. T he

Figure 1: a) (left) The diagrams show segmentation inspection progress and results. Defect size can be measured accurately to determine a “Good” or “NG” defect
based on the specification; b) (right) The diagrams show classification inspection progress and results. Different classes for each defect and multiple classes within each
defect can be classified as applicable.

31
Chip Scale Review September • October • 2020 [ChipScaleReview.com] 31