Page 36 - ChipScale_Sep-Oct_2020-digital

by bump damage classification modes to in series, product yield is enhanced a effectively enhances detection capability
differentiate true reject and pass units few percent by overkills mitigation. The and improves product yield.
(see Figure 3, right). reason this is so is because classification
Figure 4 depicts the progress of deep helps to identify the defects class Second-inspection process and
learning and combined algorithms over (mode) and judge them correctly based results
fifteen weeks of evaluation time. The on the customer specification, while One of our key achievements with
use of AI gradually improves the quality segmentation just re-checks defects respect to AI development is to integrate
by learning over the first weeks of use, detected by machine vision, to judge our machine vision and deep learning
thereby increasing the number of defects the existence of the defect, but not the in a unique way. Figure 5 shows deep
found vs. time. Starting from week 6, defect mode. Therefore, the integration learning and machine vision integration
when we apply the combined algorithms of two algorithms into our AI system as a second-inspection process. Our image
processing library includes geometric
and subjective inspection. Geometric
inspection is clear judgment by machine
vision. Examples for these inspection
modes are package XY size, bump width,
bump XY position offset, etc. In contrast,
subjective inspection involves those defect
modes for which adding AI to machine
LEADERS IN vision can significantly improve detection
capability. Examples of subjective modes
MICRO DISPENSING are FM (particle), crack, pattern/copper
exposed, etc. At the first stage in Figure 5,
TECHNOLOGY subjective inspection is done with criteria
tighter than the customer specification by
machine vision. Reject and overkill results
SMALL REPEATABLE VOLUMES are sent to the AI engine that is already
ARE A CHALLENGE, BUT NOT trained. The AI model re-inspects (second-
IMPOSSIBLE IF YOU HAVE BEEN inspection process) reject units based on
CREATING THEM AS LONG AS WE HAVE. the customer specification to differentiate
final reject and overkill (pass) units. In this
approach, deep learning is implemented as
a closed-loop feedback to vision software.
TO DO IT WELL, Therefore, the AI model distinguishes
WE PROVIDE THREE THINGS: between true reject and overkill so it
directly impacts package/product yield.
Deep-learning second-inspection
results are shown in Figure 6. We
have reviewed yield improvements
Dispensing Expertise in a variety of microelectronic of a certain product over a period of
packaging applications. one month. During the first half of the
month, a single-digit yield increase
Feasibility Testing & Process Verification based could be attained. Furthermore, the
on years of product engineering, material flow testing AI engine deepens its learning by
and software control. inspecting more units and increasing
the number of defects during this
Product Development for patented valves, period. On the 17 day of the month,
th
dispensing cartridges, needles, and accessories. when process excursion or escapee
(under-reject) has occurred, tightening
the prime inspection criteria in our AI
system (Figure 5) is an effective way
to overcome excursion and under-reject
(escapee) risks. By doing so, the yield
Our Micro Dispensing product line is proven and trusted by loss percentage jumped drastically.
manufacturers in semiconductor, electronics assembly, medical Therefore, the second-inspection
device and electro-mechanical assembly the world over. process by our AI system compensates
www.dltechnology.com. for a two-digit yield loss and significant
yield improvements are achieved every
day in the second half of the month. In
216 River Street, Haverhill, MA 01832 • P: 978.374.6451 • F: 978.372.4889 • info@dltechnology.com
this particular case, the top yield gain

34
34 Chip Scale Review September • October • 2020 [ChipScaleReview.com]