White Paper

Manufacturing | Artificial Intelligence

High-Value AI in Intel’s Semiconductor

Manufacturing Environment

With over two decades of experience in developing and applying artificial

intelligence (AI) to specific use cases, Intel has established best practices that
enable AI at scale to generate optimal business value

Authors Executive Summary

Rao Desineni
As one of the world’s leading silicon manufacturers, Intel continually strives to
Senior Director, Manufacturing, Supply
increase product quality while improving manufacturing efficiency. Artificial
Chain and Operations
intelligence (AI) is a powerful tool to transform vast amounts of manufacturing data
Eugene Tuv into insights that can improve the manufacturing process. To maximize success,
Intel Fellow, Technology Development Intel has fine-tuned its approach to AI in manufacturing by focusing on applications
that demonstrate substantial business value, practical feasibility, and quick time
to value. Each solution solves a specific, well-defined problem with measurable
Table of Contents success metrics. We’ve deployed a large range of manufacturing AI solutions
Executive Summary. . . . . . . . . . . . . . . . . . . 1 involving thousands of AI models at scale over the last 20 years. Each successful
AI solution is proliferated across all of Intel’s factories for all products.
Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . 2
Applying AI in Semiconductor Applying AI to more use cases, in more places, for more businesses, through deep
Manufacturing: Solving the n-i, n investments in an open AI platform is one of Intel’s top missions. From inline defect
Problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 detection to end-of-line yield analysis and many analytical steps in between, Intel’s
Deploying AI at Scale in Intel’s at-scale manufacturing AI solutions have generated millions of dollars in business
Manufacturing Environment. . . . . . . . . . 2 value, accelerated manufacturing processes, and contributed to higher yield and
productivity.
Prioritizing AI Use Cases to Optimize
Business Value. . . . . . . . . . . . . . . . . . . . . . . . 3
Approaching AI with a Clear Purpose
and Goal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
An AI Sampler: Three At-Scale AI
Solutions Boosting Manufacturing
Quality and Productivity. . . . . . . . . . . . . . 4
ADC with Computer Vision and
Machine Learning. . . . . . . . . . . . . . . . . . . . 4
Root-Cause Analysis. . . . . . . . . . . . . . . . . 5
Sort Test Probe Card Inspection. . . . . . 5
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Further Reading. . . . . . . . . . . . . . . . . . . . . . 6
White Paper | High-Value AI in Intel’s Semiconductor Manufacturing Environment 2

Introduction AI solutions in semiconductor manufacturing must include at

least one of the following:
Semiconductor manufacturing is complex, with tens of mask 1. Detect a problem at step n.
layers; hundreds of process steps; thousands of pieces of
2. If a problem is detected at step n, rapidly find its root
equipment, each with tens to hundreds of individual sensors;
cause using the data collected at the previous n-i steps,
and tens to thousands of integrated circuits (ICs) per wafer.
where 0< i ≥ n.
Each wafer contains billions of transistors and interconnect
lines and hundreds of electrical and physical measurements 3. Predict the outcome at step n using the data from the
per wafer, and tens of thousands of wafers are manufactured previous n-i steps, where 0< i ≥ n, and devise control
in multiple fabs every week. methods to optimize outcomes at step n.

This complexity leads to hundreds of petabytes of data

Deploying AI at Scale in Intel’s Manufacturing
during the manufacturing of products in advanced
technologies. Semiconductor companies have always been Environment
leaders in generating and analyzing data. In the artificial Some of the AI applications deployed in production in Intel
intelligence (AI) spring that we currently find ourselves in, factories include the following (also see Figure 1):
the obvious question is: Can AI be used in semiconductor
• Inline defect detection
manufacturing to gain insights from all this collected data?
• Tool/fleet/fab matching
Intel has been developing and using AI-enabling • Multivariate process control
methods — machine learning, deep learning, computer vision
• Automated wafer map pattern detection and classification1
and image processing, advanced multivariate statistics,
operations research, and others—in various aspects • Fast root-cause analysis (RCA)
of IC technology development (TD) and high-volume • Detecting outliers at sort test for both test time reduction
manufacturing (HVM) for almost two decades. The complexity and quality improvements in downstream shipped products
of TD and the scale of HVM forced us to steadily replace rule-
Behind the scenes, multiple AI-enabling techniques, such
based systems with learning-based systems (aka AI) for use
as advanced statistics, machine learning, optimization, and
cases where it made sense. And the journey continues.
various forms of computer vision are used, depending on the
use case.
Applying AI in Semiconductor Manufacturing:
Once we develop an AI solution for a specific use case and
Solving the n-i, n Problem validate its business value, we proliferate it across Intel’s
Semiconductor manufacturing comprises many steps— entire line of factories—thereby optimizing the return
process, measurement, shipping, accounting, and others. on investment for our efforts and promoting consistency
Data is generated at each of the n steps that a silicon wafer between factories.
or die must traverse before revenue is realized.

Probe Card Defect

Detection and
Classification
Automatic
Defect Classification Predictive Die Kill
and Test Time
Reduction

Wafer Map
Process Tool
Pattern
Matching
Classification

Example AI Applications
ML-based
Advanced Deployed In Intel ML-based
Process Root-Cause
Control
Manufacturing Analysis

Figure 1. The adoption of AI in Intel’s factories includes various applications.

White Paper | High-Value AI in Intel’s Semiconductor Manufacturing Environment 3

Each application in Figure 1 fits the n-i, n problem statement. Approaching AI with a Clear Purpose and Goal
Here are some examples:
The internet abounds with articles highlighting the colossal
• RCA. Data from n-i steps is used to find the root cause for failure rates of AI and big data initiatives across industries.
the abnormality observed at step n. The AI project failure rate ranges between 60 percent and 85
• Machine learning for Advanced Process Control (APC). percent, 2 despite huge investments in AI. One of the primary
Data from n-i steps is used to control the process at step n. reasons for failed AI projects is that some are created without
a clear use case in mind. In our opinion, AI should not become
• Predictive Die Kill and Test Reduction. Machine-learning a solution in search of a problem.
models are built using data from n-i steps to predict failures
at some downstream step n. Advanced Analytics Building Blocks
• Automatic Defect Classification (ADC). Applications are
devised to detect anomalies and quantify baseline non-
systematic defects at a given step n.
30%
10% Modeling
Prioritizing AI Use Cases to Optimize Understand
the Problem
Business Value
Defining a framework for prioritizing and deploying AI 10%
Solution Deployment
is critical for success. It’s important because enormous (DevOps)
volumes of data are generated from manufacturing
operations every day. Equally essential is the burgeoning 50%
interest in AI from engineers and executives, as well as the Data Management
pervasiveness of self-learning courses and the relative ease
of creating pilot AI solutions. The prioritization process
turns out to be simple, but not trivial, comprising three main
ingredients: substantial business value, feasibility, and time
to value (see Figure 2). Specifically, it is critical to answer the
following questions to assess the need to automate existing Figure 3. Building blocks of an advanced analytics solution.
business processes and workflows with AI:
• Are there any cost, productivity, or yield benefits?
• Is the targeted application error-tolerant? That is, are
occasional false positives acceptable? As well-documented on the internet, the failure rate for
AI projects ranges between 60 percent and 85 percent
• Can the solution be truly automated, at scale? That is, can
it be integrated into existing manufacturing automation across industries, despite huge investments in AI. We
systems in a way that the AI models can be built, monitored, believe that one of the primary reasons for failed AI
and updated with minimal manual intervention? projects is that a majority of them were created without
• Will the solution be ready in time to produce the intended a clear use case in mind.
business impact?

Feasibility Besides choosing a specific use case, another aspect of AI

project success is to take a holistic approach, covering four
essential ingredients: understanding the problem, solution
deployment, modeling, and data management. Ignoring one
ingredient puts the entire project at risk:
Successful AI 1. Although it represents only 10 percent of the pie-chart,
Time to
Project Elements Value it’s crucial to understand the problem to be solved. There
is no substitute for domain expertise. AI implementation
cannot be isolated from the business functions where
the solutions will be ultimately deployed. We start by
Business bringing domain experts—process engineers, equipment
Value engineers, and yield engineers—into a working group to
first understand the problem statement and associated
business value before generating a proof of concept (PoC)
Figure 2. Prioritization matrix for AI projects. AI solution. The PoC is thoroughly validated by the domain
experts and iteratively perfected prior to the next step.
White Paper | High-Value AI in Intel’s Semiconductor Manufacturing Environment 4

2. Solution deployment (DevOps) accounts for another typically maintains only 70–85 percent accuracy over time
10 percent of the advanced analytics pie. DevOps is the due to multiple reasons, including:
only way to democratize complex algorithms across • The work is highly repetitive.
the organization. Even at the PoC stage, we are already
• Process changes can result in new defect types that require
planning for HVM deployment, which includes integrating
further training.
the solution into our factory automation systems, to
ensure adoption. • Classifying IC defects is inherently difficult. Some defects
require design layout cross-referencing for accurate
3. Modeling represents 30 percent of the advanced analytics diagnosis, while others simply cannot be perceived by the
pie. Here, we follow two rules: human eye and brain.
n Start with simple interpretable techniques, such as robust
linear models or single decision trees, before employing
less transparent AI methods such as ensembles or much 90%
heavier neural nets.
Accuracy
70%-85%
n Use the best possible AI engines (algorithms) that are

Classification Accuracy
Accuracy
customized for our data domain. Specifically, we have
invested heavily in exceedingly performant engines that
can handle the unique characteristics of semiconductor
data—highly imbalanced data, missing data, categorical
data, and very often, “dirty” data. We constantly
benchmark the performance and accuracy of our custom
AI engines against the open-source engines used by
millions of data scientists around the world. 3 Depending
on the use case, we may use our custom AI engines or
the open-source ones, or a mix of both. 0 3 6 9 12 15 18 21 24
4. Data management is the biggest slice of the advanced Months
analytics pie—about 50 percent. And yet it is often the
least exciting facet of advanced analytics. The problem Figure 4. Defect measurements – accuracy vs. time.
with data is that it lives in different formats—structured
and unstructured, video files, text, and images—and is With the problem statement clearly understood, we worked
stored in different places with different security and privacy with a cross-functional team of process, yield, defect
requirements. Our data challenges are similar to those at metrology, and equipment engineers to implement a
other companies; we have addressed them and continue machine-learning (including deep neural nets) ADC solution.
to improve upon them, through multiple data initiatives. This solution has since been deployed in both TD and HVM
for every technology node manufactured at Intel—including
Intel® Xeon® Scalable processors and Intel® Optane™
An AI Sampler: Three At-Scale AI technology. The deployment itself required enormous effort
Solutions Boosting Manufacturing and investment to integrate the AI algorithms into factory
automation systems. Integration included several levels:
Quality and Productivity
• The input side with the defect inspection systems.
Thousands of AI models are in production at Intel. We used
the project prioritization framework to develop and deploy • The user side to allow defect engineers and technicians
each of the AI solutions shown in Figure 1. Let’s examine to label images and configure corresponding target layout
three AI solutions in our factories, to illustrate how the information.
prioritization framework contributes to project success. • The factory operations side to automatically generate
statistical process control (SPC) alerts and place factory
ADC with Computer Vision and Machine Learning lots on hold.

ADC with computer vision and machine learning was among ADC enables us to measure and classify most of the defects
the first AI solutions that we deployed in HVM more than a on wafers produced by Intel’s factories with a required rate
decade ago.4 Inline defect metrology helps detect excursions of accuracy, and we have not experienced any increase in
and catch issues in nanometric silicon chip layouts before the total cost of ownership compared to other solutions. We
they become serious yield and quality issues. ADC helps have also been able to use existing imaging equipment in
pinpoint issues at their source. the post-wafer manufacturing process to implement ADC
with computer vision and machine learning where it did
As illustrated in Figure 4, it can take six to nine months to not previously exist, helping to prevent errors early and
train people to manually classify defects with 90 percent increase yield at no additional cost.
accuracy. Even after training is complete, an expert operator
White Paper | High-Value AI in Intel’s Semiconductor Manufacturing Environment 5

seamlessly integrating machine-learning analytics on top of

ready-to-use fast data infrastructure, we have substantially
Applying AI to the Manufacturing of the reduced the repetitive tasks of finding, extracting, cleaning,
Latest Innovations from Intel and joining data.
Two of Intel’s flagship products are Intel® Xeon®
Scalable processors and Intel® Optane™ technology. Sort Test Probe Card Inspection
Intel Xeon Scalable processors are optimized for Another practical AI solution deployed in Intel factories
many workload types and performance levels, and happens during sort test (see Figure 5). Sort test is the final
deliver built-in AI acceleration and advanced security step in wafer manufacturing, where the individual dies on a
capabilities. Intel Optane technology—including Intel wafer are tested to determine yield (the number of good dies).
Optane persistent memory (PMem) and Intel Optane Technicians use a piece of hardware called a Probe Card to
SSDs—is transforming the memory and storage deliver test patterns to the dies on the wafer, wherein tester
hierarchy in the data center by bridging gaps, reducing pins make physical contact with the Probe Card.
bottlenecks, and crushing data latency.
The problem is due to the physical nature of the contact.
Because we deploy our AI solutions at scale across all Probe Cards are subject to wear and tear, which in turn can
Intel factories, the resulting productivity and quality confound the test results. Historically, technicians periodically
enhancements are available for these products as well used a microscope to manually inspect the Probe Cards. This
as all the other silicon products Intel manufactures. task is laborious, time-consuming, and presents a significant
ergonomic risk. Additionally, this method suffers from the
other limitations highlighted earlier in the ADC section.

Manual Inspection Automated Inspection

    Labor, Ergonomics, Quality Computer Vision, Deep Learning

Root-Cause Analysis
Another good example of at-scale machine learning and
advanced analytics is the RCA solution, which we have
democratized across all technology nodes in Intel factories.
In semiconductor manufacturing, rapidly finding the
root cause of yield and quality issues is critical for both
profitability and customer satisfaction. Probe Card damage defects
automatically detected and classified
The problem: At the scale of Intel’s manufacturing, finding the
Dust, Damage, Scuffs
root cause for a yield issue typically requires mining billions
of parameters across e-tests, SPC, tools, operations, defects,
queue times (QTimes), process times, wafer slot order, Figure 5. Probe Card defects – from manual to AI inspection.
equipment logs, and many other data types. It is similar to
finding the proverbial needle in a haystack. An analytically
We took a multi-stage approach to build a fully automated
savvy engineer, with strong domain knowledge and years of
inspection system for Probe Cards. At each stage, we built
experience, might be able to intelligently mine all available
intermediate applications that reduced the technician’s
data in a few hours or days; but this knowledge is difficult
workload. The overall solution now coalesces these
to share between even two engineers, let alone across all of
applications. For example, one application automatically
Intel’s factories.
collects the image data while the Probe Card is on the test
To democratize RCA, we developed interpretable machine- equipment. Review tools flag only anomalous areas of the
learning engines (including enhanced decision trees, novel Probe Card. Another application allows the technicians
committee methods, feature selection, and rule induction to easily label the data, which in turn allows us to create
techniques) that can handle massive, noisy, heterogeneous, a labeled dataset to train a deep-learning AI system. By
and frequently missing not at random (MNAR) manufacturing starting with a minimum viable product and augmenting the
data. These engines provide solutions for tasks such as RCA, functionality incrementally, we developed a solution that
but a large effort is required to first transform the data into worked for the technicians even as they helped us attain our
an analyzable form. We applied semiconductor domain goal of full automation. The system is now fully automated
expertise to create a custom big data storage infrastructure and deployed at multiple factories, providing significant
that provides very fast data access to the multi-dimensional productivity enhancements: A task that previously took as
data required for RCA; engineers can now find potential root long as 46 hours per week per factory has been reduced to
causes in a few minutes compared to hours or days. And by less than 60 seconds.
 White Paper | High-Value AI in Intel’s Semiconductor Manufacturing Environment 6

Conclusion Intel manufacturing is on a path to implement AI solutions

where they provide business value. We continually modernize
AI is poised to bring transformative changes to our automation systems, data infrastructure, and most
semiconductor manufacturing. We have been using a variety importantly, organizational culture, knowing that the greatest
of AI solutions in Intel factories for almost two decades now advantage comes when humans and machines work together.
and have experienced their value in terms of yield, cost, and
productivity gains. The solutions described in this paper
are examples of when it made sense to allow machines Further Reading
to do the repetitive tasks that humans would otherwise You may find the following resources helpful:
do painstakingly. In these examples, AI techniques offer
• Goodwin, Randall et al., November 2004. “Advancements
more precise results, especially when compared to work
and Applications of Statistical Learning/Data Mining in
performed by a less experienced engineer.
Semiconductor Manufacturing,” Intel Technology Journal,
However, AI is not magic. In each case, after the problem Volume 8, Issue 4.
statement was clearly understood, the AI algorithms had to • “Faster, More Accurate Defect Classification Using Machine
be selected, adopted, or developed from scratch by machine- Vision,” Eugene Tuv et al., IT@Intel white paper, November 2018.
learning experts. PoC solutions had to be extensively
• “Streamline Deep-Learning Integration into Auto Defect
validated by users and the perfected algorithms had to be
Classification,” Eugene Tuv et al., IT@Intel brief, July 2020.
integrated into factory automation systems via DevOps. Also,
while we generate hundreds of petabytes of data, it is critical • “Spatial Patterns in Sort Wafer Maps and Identifying Fab Tool
to apply AI to those use cases that can provide substantial Commonalities,” Eric R. St. Pierre; Eugene Tuv; Alexander
business value, practical feasibility, and the quickest time to Borisov, Proceedings of the 2008 IEEE/SEMI Conference
value. Once use cases have been prioritized, it is crucial to and Workshop on Advanced Semiconductor Manufacturing.
make appropriate investments compute resources, DevOps, • Isabelle Guyon et al. “Analysis of the AutoML Challenge
and integrating the algorithms into existing workflows and Series 2015-2018,” Automated Machine Learning. The
automation systems. Investments are also important to free Springer Series on Challenges in Machine Learning, (2019):
up bandwidth for domain experts. pp 177-219.
Many related trends are contributing to the ongoing AI spring
in Intel’s factories:
• Storage and compute are becoming more affordable. Find the solution that is right for your
• Human talent is increasing by virtue of abundant AI courses
and an active open-source community.
organization. Visit Global Manufacturing at
• AI momentum is growing, based on incremental success.
Intel or contact your Intel representative.
We expect the current AI spring to continue at least into
the next decade. In semiconductor manufacturing as a
whole, we are seeing growing awareness of AI across OEM The authors would like to acknowledge the contributions
suppliers, electronic design automation (EDA) vendors, data of many brilliant Intel employees from the Technology
infrastructure providers, and our competitors. Just as the Development; Manufacturing, Supply Chain & Operations;
Industrial Revolution took a while to mature, the vision of “AI and IT organizations who helped develop and deploy the
everywhere” will take time to become a reality. We are certain AI solutions discussed in this publication. Special thanks
it will happen much faster than the former, but it will take also to all those users—engineers, technicians, operators,
some enterprise-wide cultural changes to supplement the and managers—whose passionate involvement helped
technical solutions and make AI ubiquitous. bring these solutions to see light of the day.

1
“Spatial Patterns in Sort Wafer Maps and Identifying Fab Tool Commonalities,” Eric R. St. Pierre; Eugene Tuv; Alexander Borisov, Proceedings of the 2008 IEEE/SEMI Conference and
Workshop on Advanced Semiconductor Manufacturing.
2
“85% of big data projects fail, but your developers can help yours succeed,” Matt Asay, TechRepublic, November 10, 2017.
3
Isabelle Guyon et al. “Analysis of the AutoML Challenge Series 2015-2018,” Automated Machine Learning. The Springer Series on Challenges in Machine Learning, (2019): pp 177-219.
4
“Faster, More Accurate Defect Classification Using Machine Vision,” Eugene Tuv et al., IT@Intel white paper, November 2018.

Performance varies by use, configuration and other factors. Learn more at intel.com/PerformanceIndex. Performance results are based on testing as of dates shown in configurations
available updates. See configuration disclosures for details. No product or component can be absolutely secure. Your costs and results may vary. Intel technologies may require enabled
hardware, software, or service activation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the
property of others. © Intel Corporation  0222/AREY/KC/PDF  349850-001US