R&D
❯
Product Improvment
❯Improvement of productivity of semiconductor manufacturing
Leveraging AI to enhance adhesive quality
For:
Manufacturing industries; suppliers and buyers; environment
Scope:
Batch/continuous/discrete manufacturing (Deployed in 75+ manufacturing lines
in 10+ countries; specifically identify the contributors to quality; predict
potential quality failures).
Goal:
Other
Improvement of productivity of semiconductor manufacturing
For:
Executives of semiconductor manufacturing companies
Scope:
Analysis of data taken from production equipment and improvement of productivity based on the analysis.
Goal:
Other
Generative design of mechanical parts
For:
Organizations, designers, customers, end users
Scope:
Help mechanical engineers design lighter, strong, and better parts.
Optimization of ferroalloy consumption for a steel production company
For:
Steelmaking, steel industry
Scope:
Recommendation for the optimal consumption of ferroalloys by the ladle furnace
treatment during secondary steelmaking.
Goal:
Other
New machine-learning simulations reduce energy need for N95 mask fabrics
For:
Manufacturing companies of high density mask fabrics, in this project 3M specifically.
Goal:
Improved Product Development / R&D
AI solution to calculate amount of contained material from mass spectrometry measurement data
Scope:
Calculating the amount of contained material from mass spectrometry
measurement data using chromatography.
Goal:
Reduce costs
Improvement of productivity of semiconductor manufacturing
For:
Executives of semiconductor manufacturing companiesGoal:
OtherProblem addressed
Cost reduction of semiconductor manufacturing.
Scope of use case
Analysis of data taken from production equipment and improvement of productivity based on the analysis.
Description
This use case consists of the following three themes.
1. Support of analysis of cause of failure based on wafer
map patterns
At the final stage of semiconductor manufacturing, each chip
on a wafer is tested and a pattern of the failed chips are
distributed on the wafer is produced. Analysis of the cause of
failure is carried out based on the pattern and the history of
usage of manufacturing devices. The analysis is supported by
the following four technologies.
1.1 Clustering of wafer map patterns
Clustering of the wafer map patterns is carried out in order
to grasp an overview of the occurrence of the failure. Because
there are 200 thousand wafers per month, a fast clustering
algorithm is required to promptly provide information to
engineers. Making use of scalable k-means++, the clustering
process is 72,5 times faster than the previous method.
1.2 Cause estimation based on pattern mining
If failure of a particular manufacturing device frequently
occurs in the history of a wafer belonging to a wafer map
cluster and failure of the device seldom occurs in the history
of other wafers, then the device is likely to be the cause of the
failure. The candidates for the cause of the failure and their
likelihoods are calculated based on the number of
occurrences of the combinations of the devices promptly
counted by a pattern mining algorithm FPGrowth and
ranking through chi-square test.
1.3 Wafer map classification based on CNN
A wafer map is classified into registered typical wafer maps
in order to monitor the recurrence of the failure. The
classification accuracy (F1 score) with SVM was 0,898.
Making use of CNN, the accuracy is improved to 0,95.
1.4 Web portal for yield analysis
The information provided by the above technologies is
shown in a web portal. The portal has improved the average
analysis time from six hours to two hours.
2. Automatic classification of scanning electron
microscope (SEM) images of defects
Tests of wafers are carried out not only at the final stage of
the production but also between processes, where the result
of the previous processes is checked. One of the tests is
classification of images of microscopic aspects of the defects
observed by SEM. Thirty thousand images are taken daily. It
is an important test because the class of a defect may provide
valuable insight for cause estimation. Previously, the
classification was carried out semi-automatically by an
engineer with a tool with classification function. However,
the human workload was relatively high because the tools
ability was quite limited. Making use of CNN, the number of
defect categories that are automatically classifiable has
dramatically increased. Now the automation ratio is 83 %,
improved from 49 %.
3. Analysis of cause of variation of quality characteristic
value
In the Yokkaichi operation, the cause of the variation of a
quality characteristic value is identified and the yield is
maintained by countermeasures. For quick identification,
various data including process parameters and sensor
measurements from a manufacturing device are stored in a
DB. Therefore the number of attributes becomes huge by the
time of the completion of production. It is not uncommon for
the number of attributes to be much greater than the number
of products to be analysed, sometimes by several orders.
Making use of Lasso regression for data with 23 600
attributes and 303 products, a regression model predicting a
quality characteristic value has been built, with automatic
feature selection. Engineers cause identification tasks are
also supported by a network diagram that visualizes the
causal structure of the selected features. As a result, the
average analysis time is improved to one day from seven
days.
This proposal is based on the use case collection initiative
promoted by Japanese Society of Artificial Intelligence (JSAI).
Live Video
AI: Perceive
Computer Vision
AI: Understand