Semiconductor Manufacturing QA/QC using Visual AI

Demand for semiconductors is rising due to increased need for processing power. As new models are being designed faster and faster, QA/QC is becoming an expensive and time-consuming process. Visual AI with Samsung’s Autolabel can help improve QA/QC in terms of cost and efficiency.

Fast iteration

Semiconductors are critical to the manufacture of computer parts. However, there is a global shortage of semiconductors (Wu, Savov, and Mochizuki 2021). The limited number of foundries, the small handful of companies dedicated to manufacturing them, and the increasing need for semiconductors in almost all modern, digitally-enabled goods has led to a shortage of supply capacity, but a glut of demand.

The industry is primed to expand to meet this demand. The US government has recently approved 1.8 billion to expand the industry (Kelly 2021). A new factory is being opened in Germany, to help with making circuitry for the automotive industry (Busvine 2021). The industry’s expansion will help meet this demand, and enable innovations in the automotive, energy, healthcare, and other industries.

The industry has standards of quality alongside every other industry. Once ingots primarily made of silicon are forged, they are cut into thin wafers where meticulously designed circuit designs are stamped upon them. They are then cut into individual circuits, which are then used by other companies to manufacture electronics. Each of these steps needs checks and balances to ensure that the final circuit is both functional, fits the original intricate design, and contains no minor faults. This involves quality assurance and quality control practices. ?

Quality Management

Quality management is a challenge in that it requires good processes to make the product, as well as checks and controls to ensure consistency. Quality assurance is the process of ensuring quality output, often through analyzing and improving on work processes. Quality control is more active, where flawed products are removed from the assembly line actively to ensure that flawed output doesn’t end up in the supply chain. Both add time, work, labor hours, and therefore costs to the manufacturing process. These processes become hard to maintain as more products are being made, and more iterations of existing builds come online. For semiconductors, these processes are complex, time-consuming, and pricey.

Circuitry and semiconductors can be checked for flaws in the build visually; if the individual closed circuits are flawed, it will appear in how they are stamped on the semiconductor wafer. A human, using scanning technology, could review each circuit or the wafer as a whole to check for issues. Flawed product can then be removed, and then the stamps or design can be reviewed for faults.

Most manufacturers operate on a Six Sigma QA process. Semiconductors have trouble meeting these standards without removing flawed wafers and chips, resulting in significant scrap. The process can result in removing upwards of 20% of the original wafers due to flaws. This lost production can cost tremendous amounts. A single FAB unit can cost (in investments) US$3.5 billion for a five-to-eight-year cycle; this amounts to between US$87.5m to US$140m a year in losses due to removed flaws. What’s more, these losses exacerbate the critical issue of shortages.?

All of this adds to the cost of semiconductor production. This exacerbates the problem of semiconductor shortages, making the break-even cost for the chips even higher. If the supply expands rapidly due to government funding projects and new investments, there will be a need to lower costs to ensure that manufacturers can stay profitable.

Typically, QA and QC are not areas that businesses find cost savings from. These processes are necessary expenses to avoid product losses and conflicts with business partners.

But there may be a way to speed up the QC process and fine-tune it with visual AI.

Visual AI: A potential solution

Instead of having dozens of workers manually reviewing each piece, it can be done by a lone AI, who is checked and trained by these same people. A neural network could be trained with images of various permutations of the completed, pre-cut wafers to be able to identify flaws against the correct version of the semiconductor design. Visual AI can pick out the particular features of an image, such as flaws in the semiconductor print, at immense speed. The piece can then be sent to a QA specialist for review, to ensure that the manufacturing process is not impaired in some manner.

This solution is already being applied for QC in various settings. It is fairly useful, able to help minimize misprinted circuits and cut down on waste. But it is not a perfect solution. Visual presents its own set of challenges, namely the time and data required to “tech” a computer to be able to perceive and

The challenge is that the visual AI will need new training images for each different orientation of the semiconductor, and if changes are made to the design, it will need re-training. This can amount to thousands of images, meaning that there will need to be thousands of flaws logged and manually labeled to train the AI. This for every different design, even if it is minutely different. Each labeled image will require a QC specialist to identify and mark out the flaws in each superconductor print, and feed that data back into the AI before it can do the process itself.

Agility from Autolabel

Samsung SDS has a new solution to the challenge of training AI. Instead of manually labelling thousands of images, Autolabel can accomplish this process in a fraction of the time. This is because the training images are labeled not by hand, but by another AI in much less time.

By applying a gaussian classifier process to the bulk of the training images, the features of the semiconductor can be identified by the AI ahead of labeling. Next, a series of training images are manually labeled (around 200) and fed into the machine. They are then ranked in terms of how “confusing” they are. These are then manually labeled again, providing the AI with the most novel information. After a few iterations, this process trains the AI to read and identify visual characteristics in complex images with only 1/10th of the training images required.

This active learning process creates an AI that can do the job of labeling training images for an AI QA/QC device automatically. For the now growing semiconductor industry, this provides new and sophisticated quality controls, while allowing it to be agile and flexible for a diversity of designs and circuits.

Bibliography

Busvine, Douglas. 2021. “Bosch Opens German Chip Plant, Its Biggest-Ever Investment.” Reuters, June 7, 2021, sec. Technology. https://www.reuters.com/technology/bosch-opens-german-chip-plant-its-biggest-ever-investment-2021-06-07/.

Kelly, Makena. 2021. “Senate Approves Billions for US Semiconductor Manufacturing.” The Verge, June 8, 2021. https://www.theverge.com/2021/6/8/22457293/semiconductor-chip-shortage-funding-frontier-china-competition-act.

Wu, Debby, Vlad Savov, and Takashi Mochizuki. 2021. “Chip Shortage Spirals Beyond Cars to Phones and Consoles.” Bloomberg.Com, February 5, 2021. https://www.bloomberg.com/news/articles/2021-02-05/chip-shortage-spirals-beyond-cars-to-phones-and-game-consoles.

Credits: This article was written by?Ryan Cann.