Listen to this story
Recently, at the Future Ready Technology Summit in Bengaluru, Sandeep Alur, director of Microsoft Technology Centre, gave a live demo of what an assembly line or a manufacturing plant will look like. The demo involved Sandeep using Microsoft HoloLens, augmented reality or mixed reality headset, which overlayed a complete design of manufacturing setup onto the real world. This, he deemed to be the true representation of a digital twin.
Digital twin, defined simply as a real-time simulation of the real world, has recently started to gain momentum in the semiconductor world. For example, AR is integral to Intel’s worldwide manufacturing processing from maintenance and repair, enabling them to communicate remotely to troubleshoot internationally and prepare interactive training materials.
Semiconductor manufacturing is a long, and complex process. Setting up a wafer fabrication plant requires precision, clean environments, expensive equipment and time. To illustrate, it typically takes three months for leading semiconductor manufacturer GlobalFoundries to etch and fabricate silicon wafers into multilayer semiconductors. And it becomes difficult to increase production during up cycles when there is a chip shortage, since it takes years to get new factories up and running.
Advancements in AI and digital twin technology offer the potential to accelerate the chip design and production process and help manufacturers close the supply-demand gap quickly.
Revolutionising semiconductor manufacturing
Thus, ‘digital twin’ will pave the way for collaboration and radically change worker training. The lithography equipment tools are most vital to a semiconductor fab and cost up to $40 million each (for a 300mm wafer size). But, of course, the cost will shoot up as we move towards lower nodes. An important aspect of setting up any new fab is a technology transfer fee that includes specialists coming to the new facility and training the workforce on the proposed node technology. Creating a simulated environment, however, reduces the time to get the production running at any new plant.
Source: Critical Engineering
A guidepost to digital twins by NVIDIA explains that workers can be trained on expensive systems before they’re even installed using this novel technology. Once trained, workers can qualify, operate and service these machines without having to set foot in the ultra-clean rooms where they’re installed. Thus, a virtual fab will allow specialists to design and test new processes quicker and cheaper without disrupting operations at a physical plant.
Along with creating a digital copy of an entire factory, manufacturers can use AI to process data from sensors inside actual factories and find new means to route material that can reduce waste and speed operations.
At CES 2023, NVIDIA also announced a partnership with the Taiwanese electronics manufacturer Foxconn to build autonomous vehicle platforms. Foxconn will manufacture Electronic Control Units (ECUs) for cars built on the NVIDIA DRIVE Sim platform, which, in turn, is built on Omniverse and will enable automakers to design vehicle interiors and retain experiences entirely in the virtual world. Infusing AI and the metaverse, NVIDIA intends to make manufacturing smarter and more efficient.
Foxconn is set to build the second car model for Fisker Inc and make electric vehicles for Lordstown Motors Corp and Apple products.
Thus, the future of chip manufacturing and assembly is a nexus between robotics, simulation, and machine learning. The technology is also touted to be key for long-term sustainability efforts since it gives organisations a way to model and understand how to cut emissions and energy use so they can test scenarios to ultimately reach sustainability and climate goals. In fact, research from Capgemini found that 57% of organisations agree that digital twins are pivotal to improving sustainability.
Challenges of digital twins
However, there are also challenges in adopting the digital twins technology. Juan Betts, managing director of Front End Analytics, says, “The more complex the system, the more complex the AI framework, and the more data is required if using conventional [AI] techniques. Thus, training AI has typically been the principal barrier to its use.”
To ensure a reliable output, in many cases, supervised machine learning models are used where manual input is frequently necessary to label these datasets. However, as researcher Zhihan Lv discusses, the method is costly, error-prone and time-consuming, especially in a complex and dynamic manufacturing environment. Lv cites Alexopoulo, et al.’s (2020) study to point out how the digital twin’s model can itself accelerate the ML training phase by producing an appropriate training data set and automatically labelling it through a simulation toolchain, thereby reducing user participation in the training process.
Chris Rust, founder of Clear Ventures, explained elsewhere that organisations like LAM Research, Bosch (which already uses digital twins in its German semiconductor factories), and Applied Materials, are already using surrogate machine learning models that are “more accurate and up to a million times faster than traditional physics-based simulations”. He also added that technology firms like Tignis, AspenTech, and Ansys are leading the way in this field by utilising digital twins to streamline industrial operations and make AI and ML accessible to virtually any application.
In this light, Christian Mosch, general manager at Industrial Digital Twins Association (IDTA), proposes an “interoperable” approach between multiple digital twins, where data is shared across different lifecycle phases, including design, planning, construction, training, and operation, among others. In the case of semiconductor manufacturing alone, we can see how an interoperable framework between digital twin modelling and digital twin data labelling can streamline the whole process, akin to a real-world system.
Use of AI
Similarly, on the AI front, a lot of systems are already under process in various areas. For example, some common use cases in the manufacturing domain include, demand forecasting, inventory optimization, scheduling, and predictive maintenance. Shisheer Kotha, director of smart manufacturing and AI, Micron Technology, India, notes that one of the important applications has been automated defect classification, which uses image analytics with deep learning to identify the root causes in a shorter time and contributes to the yield ramp. He added, “These solutions improve quality and early detection in assembly & test operations with a combination of IoT and deep learning technologies.”