- The Conveyor
- Posts
- Can AI Reduce the Labor Gap in Chip Factories?
Can AI Reduce the Labor Gap in Chip Factories?
Abhi Rampal on chip industry problems, AI uses cases, and pushbacks.
Gowtham Ramachandran
July 30, 2024
Photo by Laura Ockel on Unsplash
The US is on the brink of a semiconductor manufacturing boom.
Since the US passed the $52B CHIPS Act in August 2022, more than 50 new projects worth over $200 billion have been announced, including massive investments from semiconductor giants such as TSMC, Intel, and Micron.
But they are all running into a major problem - they are struggling to hire people to manufacture and operate these facilities.
According to the Semiconductor Industry Association, the US semiconductor industry could face a shortage of about 67,000 workers by 2030. McKinsey predicts an even worse scenario, with a shortfall of about 300,000 engineers and 90,000 skilled technicians by 2030.
Historical semiconductor workforce and projected 2023-2030 gap. (SIA)
So… how can industry solve this labor problem? Can AI help?
I sat down with Abhi Rampal, founder and CEO of Solid State AI, to explore this topic further.
Abhi has worked with semiconductors for over two decades now, starting as a researcher during his undergrad to now founding Solid State AI, a startup that uses machine learning to help semiconductor manufacturers improve their production yield and overall equipment effectiveness.
In this piece, you’ll learn:
The top AI use cases in semiconductor manufacturing,
The use cases seeing adoption today, and
The biggest pushbacks Abhi encounters.
Let’s dive in. 👇
The Evolution of Semiconductor Supply Chains
The semiconductor industry originated in the US in the 1960s.
Companies like Texas Instruments, Fairchild Semiconductor, Intel, and IBM, operated as vertically integrated entities, controlling every aspect of chip production from design to manufacturing.
The Fairchild Semiconductor diffusion area in 1960. (The Computer History Museum Collection)
But as the complexity and cost of semiconductor fabrication increased, the industry began to fragment and specialize.
By the 1980s and 1990s, a new model emerged with the rise of "fabless" chip design companies and dedicated foundries.
This shift allowed for greater specialization, with companies focusing on either chip design or manufacturing. Taiwan Semiconductor Manufacturing Company (TSMC) pioneered the pure-play foundry model, manufacturing chips designed by other companies.
Moving to Asia
Over time, semiconductor manufacturing gradually moved to East Asia, particularly to countries like Taiwan, South Korea, Japan, and later China.
This shift occurred due to several factors:
Lower labor costs
Significant government support
The emergence of the foundry model
Development of specialized expertise and infrastructure
Employees work at Renesas Electronics Naka wafer fabrication facility in Hitachinaka, Japan. (AFP)
By the 2000s, Asia had become a dominant force in semiconductor manufacturing, and as of 2021, over 80% of the world's semiconductor manufacturing capacity was located in Asia.
The Push to Reshore
In recent years, there has been a push to bring semiconductor manufacturing back to Western countries, particularly the US and Europe, driven by factors such as national security concerns, supply chain resilience, and the strategic importance of semiconductors in emerging technologies like AI and 5G.
But the West has found it hard to get semiconductor manufacturing facilities up and running, despite initiatives like the CHIPS Act.
Abhi shared that besides the labor shortage, one of the primary reasons the West struggles is because it lacks a supply chain for upstream parts needed to manufacture the facilities.
Abhi Rampal: To build the fabs, you need to understand what materials you need and why certain materials are important. Specific types of concrete matter because you might get dust that falls, which can ruin your system. Your piping diameters, the type of pipes, the metals being used—all of that comes into consideration.
If you haven't sourced that stuff and now suddenly have to, it will take time. If you do have the knowledge, you need to build connections to the industry, which you haven't done in the last 20-30 years. Building those connections takes time.
Moreover, working methodologies are very different. How you work in the West versus in the East is very different. In South Korea, they have three shifts that run almost 24/7, seven days a week. We don't have that attitude here. That's why things can get made so fast there.
So, the know-how, the network, and the fact that they have all the supply chains built up make it much more efficient in Asia.
TSMC’s Arizona facility under construction. (TSMC)
Take TSMC’s efforts to build semiconductor manufacturing facilities in the US for instance.
TSMC is spending over $65 billion on three high-tech semiconductor manufacturing facilities in Arizona, but production at all three plants has been delayed by 1-2 years.
Thus far, TSMC has cited hiring difficulties, cultural clashes between Taiwanese management and the US workforce, struggles to find experienced workers for complex chip-making equipment, and safety concerns as the top reasons.
Top AI Use Cases in Semiconductor Manufacturing
Semiconductor manufacturing involves thousands of individual steps in creating a single chip.
Right from manufacturing a wafer, to etching patterns on it, to eventually packaging it, and the multiple intermediate quality control tests to ensure the chips meet standards.
The sheer volume and complexity of data the industry generates, coupled with the need to make increasingly complex chips and a shortage in skilled labor, makes the semiconductor manufacturing industry a prime candidate for AI adoption.
In fact, Abhi believes using AI is the only way the semiconductor manufacturing industry can meet the demands of their customers.
Abhi Rampal: The only answer to this is intelligent automation. There is no other way that we solve this because this becomes a human problem.
In fact, if you look at the manufacturing industry, by far the first industry that has included automation, and now you see the transition happening. Well, we basically are going to force ourselves to transition to intelligent automation, AI-based automation.
AI can impact labor in two ways - it can remove the need for labor by automating workflows, and it can make labor more productive by augmenting workflows.
Let’s break this down further.
AI can automate numerous tasks across the semiconductor manufacturing process, from chip design to production and quality control. The most promising use cases include:
Optimizing chip designs
Predicting equipment failures
Detecting defects in semiconductor wafers
Managing supply chains and streamlining operations
Beyond automation, AI can serve as a powerful tool to augment human capabilities and decision-making:
Accelerate discovery of new materials and processes
Monitor and adjust manufacturing parameters in real-time
Capture and spread institutional knowledge, making it easier for employees to access critical information
Provide engineers with data-driven insights to inform strategic decisions and process improvements
A recent Deloitte report went a step further to categorize how AI could impact various workflows along the value chain based on complexity of the application and ROI potential.
Source: Deloitte
According to the report, which included responses from 53 semiconductor executives, 72% of industry leaders predict that the impact of AI on the industry will be “high to transformative”.
AI Use Cases Seeing Adoption Today
Abhi identified various steps in quality control as seeing the most adoption today.
Abhi Rampal: The low-hanging fruit for AI in semiconductor manufacturing is visual inspection. AI can automate the visualization of wafers and defect detection, which traditionally required human inspection. Many fabs are already adopting AI for this purpose, as it speeds up the process and reduces the need for human labor.
Another promising area is process control, which includes tasks such as equipment calibration, in-situ monitoring, and predicting when equipment will fail.
While the semiconductor industry has traditionally relied on statistical process control and simple linear models, AI can provide more advanced, multivariate predictions.
Abhi also said that process control, which includes tasks like “equipment calibration, in-situ monitoring, and predicting when equipment will fail” is seeing adoption.
Abhi then walked me through a typical product adoption cycle at Solid State AI.
Abhi Rampal: We focus on equipment uptime, yield variance, and predicting calibration parameters. Our software, AIMS, allows engineers to build their own AI models using their data. This way, they maintain a connection with the data and understand what the AI has learned.
They use equipment data, sensor data, recipe files, and maintenance logs. All this data is amalgamated to build accurate AI models.
We start with a test project to verify predictions. This can take anywhere from two weeks to four months, depending on the complexity. Once the test is successful, we deploy the software, which can be installed on a laptop, in-house servers, or the cloud. It's designed to integrate seamlessly with existing operations without significant downtime.
Abhi also shared the most common pushback he hears from the manufacturers. Notably, he said that engineers hate the “black-box” nature of AI because they often have no idea why AI is recommending a specific solution. When AI contradicts their intuition, they find it really hard to trust the AI.
Abhi Rampal: Semiconductor manufacturing is perhaps the most advanced industry when it comes to data analytics in the manufacturing space. But engineers are accustomed to the direct connection they feel with their data and the equipment.
They lean on simple linear models to make decisions. That’s just how they operate. But when you bring in AI, it is a very hard thing for them to feel because it’s a multivariate system that takes all your data and gives you a prediction.
So if their manager asks them why they are doing what they are doing, they have no idea, so engineers don’t like that feeling.
Despite this, the complexity of manufacturing processes and the increasing volume of data are driving a shift toward AI-driven process control. AI can help manage this complexity more effectively, leading to faster production times and higher yields. As fabs continue to realize the limitations of traditional methods and the potential of AI, I expect to see a broader adoption of AI technologies in semiconductor manufacturing.
What’s next?
AI adoption in the semiconductor industry is still in its “early innings”.
According to KPMG’s Global Semiconductor Industry Outlook Survey 2024, semiconductor companies are only beginning to implement AI across various functions:
56% of companies expect to implement AI in R&D/engineering within the next two years
49% plan to use it in Marketing and sales
42% expect to implement AI in manufacturing and operations
Abhi believes AI adoption in the semiconductor manufacturing industry is inevitable, and said it will become the first physical industry to use AI at scale.
Abhi Rampal: The semiconductor industry faces a storm: rising expenses, declining profitability, and a shortage of skilled workers.
We're spending trillions to build new facilities, but we don't have enough people entering manufacturing. Semiconductor manufacturing is so sophisticated that you need at least 10 years of education plus five years of work experience to become a senior engineer. We need 60,000 to 80,000 of these engineers, and we don't know where they'll come from.
The only answer is intelligent automation.
Companies can't wait for human resources to catch up. They must automate extensively to meet their goals and timelines. Intelligent automation is the key to overcoming these challenges and achieving the necessary return on investment.
When I asked him about how he sees Solid State AI building towards this future, Abhi shared that their current product helps enter the market, but in the long term, their goal is to speed-up processes across the entire semiconductor manufacturing value chain.
Abhi Rampal: Our goal is to enable one engineer to perform the work of 10 to 15 engineers using AI.
We're creating systems that not only automate tasks but also make critical decisions, managing the increasing complexity of the manufacturing process. Our advanced AI models incorporate process physics to ensure accurate and reliable predictions, reducing errors and increasing efficiency.
By leveraging our AI expertise, we're providing the semiconductor industry with the tools it needs to navigate its challenges and build a sustainable future.
While a tenfold increase in productivity may be an ambitious target, I believe that even more modest gains could have a transformative impact on the industry's ability to address labor shortages and drive efficiencies across the entire value chain.
PS: Do you have any questions for Abhi? Reach out to me at [email protected] and I can pass it on.
Reply