New tools can boost efficiency, improve training, and reduce knowledge loss in organizations.
When all the planned fabs become operational, the semiconductor industry is likely to face a worker shortage of 100,000 each in the U.S. and Europe, and more than 200,000 in Asia-Pacific, according to a McKinsey report. Since the dawn of technology, people have worried that robots, automation, and AI will steal their jobs, but these tools also can be put to use to help fill the chip industry talent gap by replacing humans.
There are many ways that AI, digital twins, virtual and augmented reality, and robots/cobots (collaborative robots) can help fill the talent gap. Among them:
“When the internet came along, people said, ‘Oh, it’s going to take people’s jobs from here and there,’” said Mark da Silva, senior director of the Smart Manufacturing Initiative at SEMI. “But it actually had a multiplying effect on the total labor force. The same thing is going to happen with automation in fabs. You need a whole stream of people to build the robots and to program them, so jobs will be created in the long term. Some people use the term ‘lights-out factories,’ but I don’t think that’s a realistic vision. You’re always going to need some humans in the loop.”
Every paradigm change in technology offers benefits and presents challenges. “In terms of the semiconductor manufacturing base, big companies such as Intel, Samsung, and TSMC are at the leading edge of smart manufacturing,” da Silva said. “They seem to be moving down this path of incorporating automated material handling systems (AMHS), cobots, digital twins, AI, and other technology elements in their fabs.”
On the other hand, some companies still produce chips on more mature nodes, such as 90nm or 130nm, making unique components for the automotive sector or other industries. “These organizations often use older equipment that may not fully support smart manufacturing technologies,” da Silva noted. “Given today’s wide geographic distribution of the chip manufacturing base, it’s unlikely that we will see the whole ecosystem converging to autonomous factories at the same time.”
Some older fabs also are concerned about job displacement from deploying autonomous systems, combined with the fact that their operators might not be willing to even meet with an automation vendor. Another big challenge is retraining and up-skilling workers, depending on the skill level of the engineer or technician.
Researchers at MIT, the Productivity Institute, and IBM created an AI task automation model to more accurately predict the pace of automation, looking specifically at computer vision models. They found that only 23% of vision tasks would be attractive to automate at today’s costs because the required AI systems would be too expensive to build to ensure a proper return on investment. They also noted the difference between full task automation, which is more likely to displace workers, and partial automation, which likely improves worker productivity.
Data processing, legacy record keeping in the fab
Companies that look at AI as a replacement for human staffing are thinking about it the wrong way, said David Park, vice president of marketing at Tignis, who authored a blog on how AI and ML can help beat the skill shortage. “It is a very powerful support capability for the existing engineers in the fab and helps solve problems more quickly.”
One of the main challenges AI/Ml can solve is what to do with the terabytes of data coming out of every tool every day. “About 70 new fabs are under construction around the world, and finding technicians and engineers to support all their operations will be challenging,” said SEMI’s da Silva. “Smart manufacturing technologies focused on effectively using available data, and extracting value through actionable inferences, will be needed because there’s too much for a single human — or even a team — to analyze effectively in needed timescales.”
But in terms of what a closed-loop, decision-making process should be, AI can’t do it by itself, said Park. You need the human insight to help it figure out the best resolution. Likewise, humans need to label data so AI knows what to exclude. “For example, you don’t want the AI to use data from test chips, because they aren’t representative of actual high-volume manufacturing,” he said.
Ideally, the software worlds and physical worlds should come together to solve challenges. “There is institutional knowledge that needs to be transferred from the people in the fab to the data science team and the AI/ML tools to get the maximum value out of them,” he said. “You can’t do one without the other.”
In other words, fabs need a mix of people who are book smart and street smart. And while senior hardware engineers can learn new software, they aren’t likely to become expert coders overnight. “However, there are ways that the process engineers can make it easier for hardware engineers and technicians to extract the power of AI without having to become data science experts,” he said.
To this point, Tignis offers a digital twin query language (DTQL) system, which is easily readable by non-programmers and has inquisitive, reporting, and monitoring modes, making it easy to preserve the institutional knowledge of tenured engineers who will eventually retire.
“Senior workers have determined the rules over time that are important for the fab, and it all gets documented with the rules or signals an AI algorithm is looking at, how it works, and the decisions that it’s making so people can learn in an offline mode and come back and put in comments,” said Park. “So if a rule gets triggered 10 years in the future, and the person who originally identified this issue is long gone, the current engineer can see what that person was looking for and trying to do. It’s a living document that is active within the system, and it adds value to AI so it’s not just some amorphous algorithm. You can look into it and make very subtle or substantial changes based on your personal knowledge.”
From an onboarding perspective, the DTQL can provide a visual and interactive environment where an engineer can understand the goals of the manufacturing processes, how they work, and how to troubleshoot them.
Virtual cleanroom training for engineering students
Increasingly, digital twins and virtual twins also are being used to train engineering students at universities and future technicians at community colleges. For example, Purdue University partnered with Dassault Systèmes to accelerate semiconductor workforce development through the use of 3D virtual twins, simulation, and AR.
“The semiconductor manufacturing process is very lucrative, but it’s also very obtrusive and difficult,” said Bill DeVries, vice president of industry transformation and customer success at Dassault Systèmes. “You’re in cleanrooms, all bound up in bone-dry conditions. Sometimes you’re sub-cryo. You’re dealing with dangerous chemicals. It’s hard to get people into this environment. That’s the first barrier to workforce development.”
The second barrier is the capital required to build facilities for manufacturing, R&D, or training purposes. “Purdue is putting $100 million into their cleanroom facility,” said DeVries. “That’s a significant outlay, but they can only put 10 students at any one time into the cleanroom. If we’re going to produce enough workers to repatriate chip manufacturing back to the U.S., we need far greater capacity.”
Dassault took a 3D scan of Purdue’s entire clean room, including equipment that had been ordered but not yet installed. “A digital twin is a snapshot of something digitally, whereas a virtual twin is something that lives and persists in its totality going forward,” said DeVries. “From a student’s perspective, they’re experiencing what the cleanroom looks like, what etching may look like, and every step of the manufacturing process before they get their experience in the fab, because the space in there is so precious.”
Fig. 1: A virtual cleanroom compared to the actual room. Source: Dassault Systèmes
By using the virtual twin, students learn faster and they’re more productive in the actual cleanroom because they’re already aware of their surroundings. Currently, students with a login can access the virtual twin on a laptop, tablet, phone, or PC, and Dassault is working with Apple and Meta on immersive goggles.
“You could view it in goggles today, but wouldn’t be able to interact with it,” said DeVries. “We want to integrate how people interact with the platform so they can build, and turn things on and off. There are some very good reasons why you want an immersive experience, but also practical reasons why a screen can be better. For example, if you have a smaller room compared to what the actual one is, you don’t want to be walking into a wall.”
The 3D experience platform is being used by the U.S. military and government, as well as in other countries, so it has the highest levels of data security and encryption. “The models can be turned off at the data level, and there’s a filter where you can make sure people can only see the data that they’re supposed to and lock them out of other data,” said DeVries.
Dassault helps Purdue develop the curriculum for the university’s semiconductor degrees, and there are classes that enable students to get certified on the 3D experience platform to build a cleanroom virtual twin or use the technology for other applications. The university also offers a free online course, Semiconductor Fabrication 101.
“Overall, the impact of digital twins on the workforce promises to be overwhelmingly positive,” said SEMI’s da Silva. “Digital twins offer many potential benefits, such as immersive training, real-time skills assessment, objective evaluation, improved decision-making, and enhanced collaboration. This makes it easier for geographically dispersed teams to collaborate more effectively in sharing insights, conducting experiments, and making informed decisions — all in a virtual environment. A good analogy is flight simulators to train pilots before putting them in an actual plane. Likewise, you can train technicians and engineers before they work in a fab.”
AI to support cybersecurity, development teams
Because of pandemic-era budget restrictions, rising interest rates, and reduced operating capital, companies have prioritized how they meet minimum requirements for security, said Jamie Boote, associate principal security consultant at Synopsys’ Software Integrity Group. “As firms are making these tradeoffs, they have to decide if it’s okay to use automated tooling versus experts who have a good understanding of what’s going on.”
While automating processes may look like a reduction in human headcount, it ends up allowing teams to do more with the same amount of people. “It’s a multiplier where it allows teams to scale, because we are still years away from having enough security everywhere that’s needed,” said Boote. “Many companies have just enough security to avoid a breach and stay out of the headlines, so AI is going to allow firms to go beyond that.”
For example, Synopsys’ Building Security In Maturity Model report, which analyzed the software security practices across 130 organizations, found that daily code review increased by about 68% when automated. It used to take 8 to 12 hours for a person to review an entire codebase, whereas now you can get similar checks with a targeted focus in minutes or less, according to Boote. “But you still need people to deal with the output and make sure developers aren’t being hit with so many alerts that they ignore them.”
Engineering teams also need to weed out false positives when it comes to automated security alerts in low-code and no-code software building tools. “Because environments change, technology changes. The source code is written by other humans, so the data could potentially change, and the security tools could alert you to something that isn’t a real vulnerability,” he explained. “At that interface where the tools provide the results, you need to make sure the true positives are passed on to maintain that signal-to-noise ratio. Information also needs to be routed to someone who can read the output and go, ‘Hey, we’re storing data the wrong way. That could lead to secrets being leaked, or information disclosure. We need to pump the brakes and do it the right way.’”
Overall, humans are still best at being creative, solving problems, and building things, such as designing a new application to meet a client’s need. “People can look at software blueprints and identify gaps and missing controls, or places where you need encryption or software validation. They can also look at a design and make sure the intent is followed. For example, with a payment system, AI might give the benefit of the doubt and allow payments to go both ways, but a human would know that shouldn’t happen,” Boote said.
Similarly, ChatGPT and other copilot tools can write code, but it’s not always functional, optimized, ideal, or safe. “People who are looking at AI as a way to save headcount are probably going to see security issues in the future,” he said. “You wouldn’t trust a million-dollar value stream to someone fresh out of college, and that’s what you’re doing if you’re relying solely on AI as your primary source of productivity and security. What AI can do is allow more senior engineers to supervise a team of virtual developers. But that’s a short-term fix, which will work until the senior folks have to retire. And if we’ve replaced all the entry-level coders with AI, who’s going to be the next generation of senior developers?”
Making manufacturing more appealing to younger workers
In reports concerning the benefits of automation versus legacy equipment by Visual Components, a smart manufacturing company, 42% of surveyed U.S. manufacturers and 34% of surveyed U.K. manufacturers said that hiring new talent is one of their biggest challenges. Meanwhile, businesses expect 23% of their workforce to leave over the next five years due to retirements or switching industries. More than half said they do not have a solution to deal with lost knowledge when skilled professionals leave or retire, and many have not yet built relationships with educational institutions to build a talent pipeline.
“If we think about the manual welding process, many young people are not interested in this kind of work,” said Mikko Urho, CEO of Visual Components. “In manufacturing, the study shows we need to invest on the automation side to attract workers, and also upskill the existing workforce. When you use technologies like simulation or offline programming of robots, those are like games in the manufacturing environment, so this kind of work appeals to younger people. When it’s not all legacy equipment and you also have these software programs, it makes it easier for schools and universities to recruit students.”
The previously mentioned McKinsey report also stated that software and automation help increase employee satisfaction. “Proximity of a company’s business model to software is a contributing factor. For example, workers in foundry, materials, and outsourced semiconductor assembly and test (OSAT) positions report low scores, while those in intellectual property, electronic design automation (EDA), and fabless have the highest employee satisfaction scores.”
At the same time, Ansys recently launched a virtual assistant to make simulation and modeling more attractive to young engineers. AnsysGPT uses well-sourced Ansys public data to answer technical questions about its products, combining the domain expertise normally distributed across multiple engineers into one tool.
“Simulation and modeling help to remove errors and flaws in the design process, and to run your production better so you get a return on investment sooner,” said Urho. “That will continue to grow even as the number of people on the factory floor decreases.”
Cobots
Another development in manufacturing is the use of cobots, which are industrial robots with six-axis arms and kinematics, allowing people to safely work beside them. If cobots touch something, they immediately stop, whereas traditional industrial robots require safe defenses and no one is supposed to work close to them.
“Cobots can do the automated processes, then the human workforce can do more highly skilled simulation and offline programming — and sometimes customization is easier and better with human workers,” said Urho. “Cobots don’t make humans redundant. They put them in a different hierarchy so the work is less repetitive and more attractive. You program your cobot offline. Then, when the product changes, you have the program ready and the downtime is low.”
Fig. 2: Modeling software for a welding cobot. Source: Visual Components
In some companies, senior manual welders who used to do the physical work are now running the cobot systems, and they make the best offline programmers because they know the process well, said Urho. “This generation is retiring and they have a lot of silent knowledge about the processes that is not documented anywhere if the company doesn’t use software to record it. Technologies like simulation and offline programming are also good tools that you can use to share knowledge with younger generations.”
Conclusion: Beyond Industry 4.0
The notion of a fully-automated Industry 4.0 factory is evolving into a future vision in which people hold more valuable positions in manufacturing supply chains. A European Commission report on Industry 5.0 said automation should provide people with better jobs, not make them redundant: “Industry 5.0 therefore requires social innovation to enhance prosperity and foster good quality jobs alongside measures to support education and skill training to enable workers to adapt to a shifting job market.”
Automation might only fill a fraction of the talent gap, but it will add value to a company by allowing humans to do tasks that there wasn’t time or budget for previously. AI/ML models can also add value to a company as entities in themselves.
“Machine learning models are — outside of your people — the only appreciating asset in your fab,” Tignis’ Park said. “The physical machines are depreciating, and eventually they’re going to break down and need to be replaced. AI/ML is one of those things that grows in value as an asset, especially when you take into perspective that the models can be retrained and continue to learn. They don’t stagnate, they don’t depreciate, because the automated retraining continues to adapt and grow just like your human assets in your favor.”
Related Reading
Smart Manufacturing Makes Gains In Chip Industry
Lights-out manufacturing is increasingly sought and used by fabs and OSATs.
Fabs Begin Ramping Up Machine Learning
New models can debug processes and boost yield, but there are lots of caveats.
Chip Ecosystem Apprenticeships Help Close The Talent Gap
Registered apprenticeships, community colleges, and in-house training can help build a broad workforce pipeline for technician roles and more advanced positions.
Visa Shakeup On Tap To Help Solve Worker Shortage
Adjustments to H-1B visa program could help keep highly qualified engineers in the U.S.
Chip Industry Talent Shortage Drives Academic Partnerships
Universities, companies, and governments are forming broad partnerships to update skills and foster innovation in chips, security, AI, and related fields.
 |
 |
 |
 |
 |
 |
 |
 |
 |
Leave a Reply