SK hynix Deploys Gauss Labs’s AI-Based Virtual Metrology Solution to Predict Wafer Manufacturing Process Outcomes

• Gauss Labs, an industrial AI startup SK hynix invested in, delivers an AI-based virtual metrology solution called “Panoptes VM” for high-volume manufacturing
• By predicting wafer process outcomes based on sensor data, Panoptes VM reduces process variability by 21.5% on average and ultimately improves the yield as well
• Panoptes VM has been utilized in SK hynix’s thin film deposition process since December 2022 and is expected to expand into other processes in the future

Looking to improve operational efficiency and yield in its semiconductor manufacturing process, SK hynix has turned to an artificial intelligence (AI) solution. Gauss Labs, an industrial AI startup SK hynix invested in, launched an AI-based virtual metrology (VM) solution software product called Panoptes VM in November 2022. Right after in December 2022, SK hynix began using Panoptes VM in its mass production fabs.

Panoptes VM predicts manufacturing process outcomes using sensor data. The product is named after Panoptes in Greek mythology, the all-seeing giant with 100 eyes. Accordingly, Panoptes VM is designed to monitor everything that happens during the manufacturing process.

Panoptes VM was first applied to thin film vapor deposition¹, a crucial process that coats a thin film on a wafer. The thickness and refractive index of the thin film are key process outcomes that are directly related to the quality of a semiconductor chip. However, measuring these process outcomes for such a thin film would take a great deal of time and resources, so it is infeasible to make measurements for all wafers.

SK hynix now relies on Panoptes VM to resolve this problem. Combining prediction values generated by Panoptes VM with APC², SK hynix reduced process variability³ by 21.5% on average, which also led to improvement in yield. SK hynix and Gauss Labs are considering plans to expand this technology to various processes beyond thin film vapor deposition.

¹Thin Film Vapor Deposition: A process that coats the top of a wafer by forming a thin film — a very fine layer of material on a substrate surface, e.g., insulated semiconductor, glass, and ceramic — via a physical or chemical reaction such as vacuum deposition or sputtering.

²APC (Advanced Process Control): A solution that finds optimal process conditions for equipment during a manufacturing process.

³Process Variability: The amount of fluctuation in the quality of manufactured products in respective processes. As the probability of defects falls when the fluctuation decreases, the process variability should be managed within a certain level.

Analyzing the real sensor data with AI technology, prediction models by Panoptes VM achieve a high level of accuracy comparable to physical metrology equipment. Consequently, virtual metrology allows manufacturers to monitor essentially all wafers and opens up endless possibilities through predicted values.

Mike Kim, CEO of Gauss Labs, said: “Gauss Labs is solving the most challenging problems in manufacturing by using state-of-the-art AI technology and creating real impact and values in practice. With Panoptes VM at the forefront, we will continue to develop products that will lead innovation in manufacturing.”

Regarding the adoption of Panoptes VM, Young-sik Kim, EVP of Manufacturing/Technology at SK hynix, said: “SK hynix is making concerted efforts with Gauss Labs to realize smart factories with a next level of intelligence. We will continue to maintain our technological edge by incorporating AI technology into all stages of semiconductor manufacturing. The arrival of Panoptes VM is just the beginning.”

Interview with the People Behind Panoptes VM

AI is considered one of the main drivers of the 4th Industrial Revolution, but it’s not widely known how AI is actually applied in the industry. This is due to the difficulty of commercializing and adopting industrial AI solutions in practice. Yet, Gauss Labs and SK hynix managed to introduce AI to semiconductors, one of the most complex manufacturing sectors.

Dong Kyun Yim is a project manager at Gauss Labs who has led the launch of Panoptes VM. Doh-hyung Noh is a technical leader in the AI/DA Solution Development Team at SK hynix who has ensured that Panoptes VM will be applied effectively in real production. Hyeon-kyeong Jeong is a technical leader in the Thin Film Technology Strategy Team at SK hynix, who has been working to improve the thin film deposition process by linking Panoptes VM to APC.

Q. We heard that Gauss Labs specializes in industrial AI solutions. What is industrial AI and what role does it play?

▲ Dong Kyun Yim is explaining the effect and value of industrial AI.

 

Yim “Industrial AI literally means analyzing and utilizing data in industrial settings using AI technology. Gauss Labs believes that AI has the biggest impact in the following five areas of manufacturing: “yield & quality management”, “planning, scheduling & dispatching”, “process & equipment control”, “equipment maintenance” and “process monitoring”.

“Consumer-oriented AI services such as customer service chatbots and product recommendations are already familiar to most people. However, industrial AI is still an esoteric concept to many as it takes a longer time to be commercialized due to higher accuracy and reliability requirements compared to consumer-oriented AI. For example, it’s quite obvious that inaccurate analysis of a manufacturing process will be much more damaging than an inaccurate recommendation of a consumer product.

“Industrial AI has yet to reach the commercialization level of consumer AI, which also means that it can be considered a blue ocean with endless opportunities. Gauss Labs has been focusing on products for manufacturing process monitoring and recently launched Panoptes VM, which has been deployed by SK hynix since December 2022.”

Q. What is virtual metrology?

▲ Doh-hyung Noh is explaining the concept and necessity of virtual metrology.

 

Noh “Metrology refers to the measurement of a product’s quality during a manufacturing process. But since it takes an enormous amount of time and resources to measure process outcomes, sampling is typical in high-volume manufacturing. In other words, only a fraction of the products is measured instead of measuring them all. Virtual metrology spares physical measurements and, instead, predicts the quality of products that have not been sampled by leveraging machine-generated data such as equipment sensor data.”

Jeong “More concretely, Panoptes VM for thin film deposition predicts process outcomes such as the refractive index and thickness of the film deposited on the wafer by analyzing data such as pressure, temperature, distance, amount of gas, and electric current inside the equipment chamber during the thin film deposition process. Achieving the effect of a complete measurement without actually conducting one is the advantage of using Panoptes VM.

“In semiconductor manufacturing, the demand for metrology is increasing more and more as product patterns get miniaturized. Nonetheless, with a limit on increasing physical metrology, the need for virtual metrology is growing. The very concept of virtual metrology has existed for some time, with numerous attempts to introduce it in real production.”

▲ Taken from a promotional video of Panoptes VM (Source: Gauss Labs’ official YouTube channel)

 

Q. Do you mean that there were many challenges to putting virtual metrology into practice?

▲ Hyeon-kyeong Jeong is discussing the technical challenges of incorporating virtual metrology for the thin film deposition process.

 

Jeong “First of all, it was not easy to analyze process outcomes because there were not a lot of actual metrology data. In addition, it required a huge effort to find meaningful parameters among hundreds of various sensor data that affected the process outcomes. In short, it was difficult to create a virtual metrology model with high accuracy.”

Noh “There were also many challenges caused by data drift and shift⁴. Data drift occurs when the state of equipment changes over time, and data shift stems from equipment maintenance. These effects can lead to an additional workload on process engineers who have to update the model continuously to account for changes in the data itself.”

⁴Data drift and shift: Gradual and sudden changes in the statistical distribution of the data.

Q. How does Panoptes VM overcome these challenges?

▲ Hyeon-kyeong Jeong, Dong Kyun Yim, Doh-hyung Noh, and Simon Zabrocki (from Gauss Labs) are discussing Panoptes VM.

 

Yim “Panoptes VM collects data from multiple machines performing the same process and recommends important parameters for predicting process outcomes using AI technology. It also allows engineers to customize prediction models based on their domain knowledge⁵. This combination of AI algorithms and human knowledge results in an even more elaborate selection of parameters necessary for predicting process outcomes. This naturally leads to high accuracy.

“Moreover, Panoptes VM offers an automatic model update function that uses AI to constantly learn data trends to detect and incorporate changes such as data drift and shift in real time. This relieves engineers from the need to constantly monitor and update the model.”

⁵Domain knowledge: Expertise in a specific field.

Q. What are the expected benefits of Panoptes VM?

Jeong: “When the virtual measurement results from Panoptes VM are connected to process control⁶ such as APC, the process variability decreases and the yield increases. This is because the process can be controlled even at the single-wafer level. In fact, we witnessed the process variability decrease by 21.5% on average as a result of the early introduction of Panoptes VM to main process steps, which was followed by a corresponding yield improvement.

“Additionally, we can prevent quality control accidents. In the past, it used to take a long time to identify issues in a process and take proper actions because metrology sampling cycles were long. With Panoptes VM predicting process outcomes for every wafer, engineers can now detect and resolve anomalies in process and equipment rapidly.”

⁶Process control: The ability to monitor and adjust a process to obtain a desired output.

Q. How did your work change after Panoptes VM was introduced?

▲ (From left) Hyeon-kyeong Jeong, Doh-hyung Noh, and Dong Kyun Yim are sharing experiences of using Panoptes VM.

 

Jeong “The ability to analyze data has improved for engineers while the workload of analysis itself has decreased. In the past, we could obtain only a few process outcomes from sampling, but we can now secure comprehensive all-wafer data for analysis. It is easier to grasp the correlation between equipment sensor data and process outcomes, and the insights thus obtained can be applied back to Panoptes VM to create a virtuous cycle of growth.”

Q. What is your plan following the introduction of Panoptes VM?

Noh “We plan to expand process areas to adopt Panoptes VM. Currently, it’s only applied to the thin film process, but tests are already underway for other processes. We will continue to work closely with Gauss Labs to advance into the mass production stage.”

Yim “In addition to Panoptes VM, Gauss Labs is also developing Panoptes RCA (Root Cause Analysis) and Panoptes IM (Image Metrology). Panoptes RCA identifies true causes of abnormalities during the manufacturing process and will be released early this year. Panoptes IM measures process outcomes from various images and will be released later this year. Gauss Labs will stay committed to helping field engineers with AI technology.”