What is a Semiconductor Digital Twin?

The Biden administration recently announced a $285 million funding opportunity for using digital twins in the American semiconductor industry. This follows similar investment initiatives in Japan and India for semiconductor manufacturing. In the U.S., the money will be used to fund a digital twin institute, and the money given by the government is expected to be matched by members of the institute.

While the broader manufacturing industry has already started to use digital twins, it is still a relatively new technology for semiconductor chip manufacturing. The recent White House initiative aims to build a stronger ecosystem across America, in which chip manufacturers will use digital twins to reduce reliance on manufacturing processes from other countries and build a more robust supply chain.

A digital twin is a virtual representation or model of a real-world entity or process that can be managed in real time. The virtual representation is connected to the physical asset via sensor streams and becomes a digital twin, in which you can analyze the past, present, and future behavior of a system to better understand how to optimize it.

A digital twin is built using physical sensor data that keeps the virtual environment up to date and enables the digital twin to experience the same things as the physical asset to predict in real time how it will behave.

Multiple layers and technologies are required to build a digital twin: 

• Infrastructure layers, including sensor networks that physically attach to semiconductor manufacturing equipment
• Communication infrastructure
• Computer infrastructure
• Data storage infrastructure
• Application enablement layers

Digital twins also use the Internet of Things (IoT) and edge platforms — like those available from Microsoft, NVIDIA, and Amazon Web Services (AWS) — to build virtual models of physical systems. 

The final layer is analytics. In this layer, all the collected data is combined with engineering insights (such as those obtained from simulations) to build the virtual representation or model that can then be used to make predictions about the system and provide insight into its workings. Overall, digital twins enable better decision-making when it comes to determining the optimizations that should be made to the physical system.

The semiconductor manufacturing process is highly specialized and globalized. For instance, two foundries (TSMC and Samsung Foundries) account for about 70% of all chips manufactured globally, and only one company (ASML) manufactures nearly all of the extreme ultraviolet (EUV) lithography machines that are critical to advanced node manufacturing processes. 

This reliance on a few key players in the market could potentially lead to supply chain bottlenecks like we saw during COVID-19 lockdowns. With more than 60% of chip manufacturing based in Taiwan, there is also the potential for chip shortages due to geopolitical tensions. 

The funding opportunity from the White House could alleviate some of these supply concerns by developing more semiconductor fabrication plants (known as fabs) in the U.S. While this will start with academic institutes performing the required validation processes on any new digital twin solutions, the end goal is to see more public-private partnerships through companies such as Intel, which is already funding semiconductor workforce training and education in the U.S.

Digital twins benefit semiconductor manufacturing processes by:

• Improving the yield of the process
• Bringing equipment online faster
• Providing predictive maintenance and smart manufacturing capabilities
• Reducing maintenance costs and unplanned downtime 

The ability to perform this level of optimization has been confined to a few key chipmakers, but digital twins can make this know-how more accessible to more regions of the world.

While digital twins can be used to design and prototype new chips, the main benefits will be seen in manufacturing and operations. Digital twins will help improve the process’s output by creating a virtual model of either the individual equipment or the larger supply chain environment. 

Consider HVAC and airborne molecular contamination (AMC) filter systems that prevent contamination in the sensitive chip manufacturing process. If filters are not changed at the right time, the quality of the chip will suffer, leading to expensive redos. If the filters are changed too soon, the fab will face unnecessary and expensive downtime. Digital twins offer a way to maintain filter and HVAC systems more efficiently.

Other aspects that can be modeled and optimized include:

• The gas flow network
• Components inside each piece of manufacturing equipment
• The etching and soldering processes
• The temperature at which the wafer is operated

Once the chip has been manufactured, the virtual environment can be used to check that it’s operating efficiently.

One crucial area in which digital twins add value is in their use with virtual sensors. Digital twins provide access to data that wouldn’t otherwise be available with physical sensors alone. 

You can use physical sensors at accessible points of interest and then use the algorithms to “virtually sense” or simulate the rest of the environment that might not be accessible with physical sensors — so long as the available physical data is validated.

One example is ensuring that furnaces (such as those in PECVDs) are operating properly by virtually sensing internal temperature. The temperature of the wafer needs to be strictly maintained in order to get a good yield. Physically sensing temperature on the surface of the wafer is not easy because it would affect the manufacturing process. However, using a digital twin, virtual sensors can sense and maintain optimal temperature during production. 

Since the underlying models used in these digital twins are based on physics, they can make accurate predictions for a fairly large range of operation. Additionally, statistical calibration techniques such as Bayesian calibration can improve the accuracy of the digital twins — for example, only a 1-2 °C error in a typical 1,200-2,000 °C furnace. 

Machine learning and neural network algorithms augment the simulation and improve its quality to ensure that the virtual sensor is as close as possible to the physical sensor.

So far, there has been low adoption of digital twins for semiconductor manufacturing. This is partly due to the difficulty of modeling complex nonlinear physics. However, with advances in simulation technology, several of the critical subsystems within semiconductor manufacturing equipment can be modeled with state-of-the-art technology. The bigger challenge has been in model availability. This is because equipment manufacturers will often have detailed models and domain knowledge about the equipment, but the foundries who want to use the equipment don’t have access to this information.

Digital twins could provide a mechanism to enable knowledge sharing, allowing equipment manufacturers to better understand how an operator uses the equipment and vice versa. Potential IP concerns can be addressed via restricting or limiting access to information provided by the digital twin.

Ansys focuses on taking existing simulations that customers have and converting them into a form suitable for semiconductor manufacturing, then plugging them into an IoT stack or edge computing. This approach is called reduced-order modeling (ROM), which takes complex simulations or existing simulations and converts them into a real-time model. 

Ansys software can also work with measurement data or instrument diagrams. It’s a start-to-finish process that validates each stage of the digital twin build before deploying.

Ansys has two software packages that are used in tandem to build digital twin models — Ansys Twin Builder software and Ansys TwinAI software. You can combine them with other simulation software packages for a more robust simulation environment. 

Twin Builder software focuses on the simulation and physics side of the digital twin, while TwinAI software combines the simulation with data to produce accurate evolving digital twins. Machine learning improves the accuracy of the models and ensures that the digital twin can self-calibrate to the changing behavior of equipment as it ages. Once the two tools have generated the digital twin, you can export it using containers, Python apps, or web apps. Thanks to prebuilt connectors to Microsoft and AWS digital twin platforms, it’s easy to deploy at scale.

With the White House initiative reducing the barriers to semiconductor manufacturing, now is the time to act. Much of the technology to scale already exists because other manufacturing industries have already adopted it. It won’t be long before we start to see digital twins being rolled out on a large scale across the global semiconductor industry.

Related Resources