When will AI usher in a new era of manufacturing?

Manufacturing processes are industry dependent, and even within a sector, they often differ from one company to another. However, some things are common to virtually all types of manufacturing: expensive equipment and trained human operators are always required, and both the machinery and the people need to be deployed in an optimal manner to keep costs down. Moreover, lowering costs is not the only way manufacturers gain a competitive advantage. They also get ahead by reducing production errors.

Companies across a multitude of industries are now using AI to improve their manufacturing processes. Most of the time, custom AI solutions are needed to fit into the processes of the company. Experience shows that cross-organizational teams need to work together. At a minimum, product and engineering teams need to be involved to build a solution tailored to the unique ways the company takes raw materials and components, and converts them into goods they can sell.

AI improves diaper manufacturing

“All areas of P&G’s business are being impacted by emerging technologies like automation, AI, and machine learning,” says Vittorio Cretella, CIO of Procter & Gamble. “Given this, we’re flexing our digital muscle across the enterprise and doubling down on AI to generate additional benefits for our business and our consumers. One area where we’re progressing fast is in our manufacturing environment, where digital solutions can help improve quality and resilience.”

While most of us have given very little thought to how diapers are manufactured, Procter & Gamble has spent decades trying to perfect the process. Technology has always been an essential enabler of that mission—and more recently, AI has played a critical role in improving the production of Pampers, one of the best-known brands in the world.

Diapers are produced by assembling many layers of material with great precision to ensure optimal absorbency, leak protection, and comfort. The manufacturing process for Pampers requires more than 40 separate glue streams to assemble each diaper. The glue streams must function with flawless consistency at a rate of 1,200 diapers per minute, and all this happens on 140 different manufacturing lines globally.

P&G engineers developed a high-speed data collection system to capture data to use for training AI models. One challenge they faced is that, while production errors are extremely costly and disruptive, they don’t happen often, which means that failure events are underrepresented in the training data. To fill this gap, engineers created a high-fidelity, hot-glue data simulation model to generate extra data, mimicking both a manufacturing line producing diapers with no glue failures, and a manufacturing line that fails in different ways.

Using an AI platform developed in partnership with Microsoft, P&G engineers created a model they trained using both real and synthetic data. With this model now in place on live manufacturing lines, line operators receive real-time alerts that enable them to quickly direct and address failures of manufacturing and glue stream models.

“This new model has helped us maintain the integrity and product superiority that parents have come to expect from Pampers.” says Cretella.  “From a business perspective, we’ve been able to reduce glue-related scrap by 80%, which is a very successful outcome.”

AI helps manufacture semiconductors

Not only do chip manufacturers develop some of the technology that underpins AI, but they also apply the same technology to their own manufacturing processes. “We deployed AI to do predictive analysis based on data—and I see it being deployed across the industry,” says Mark Papermaster, CTO of AMD.

AMD is fabless, meaning the company manufactures devices through partners. It also designs semiconductors, which are manufactured by foundries. Then they package the components to create the final product. A massive amount of data is already collected from sensors across all processes and from all supply chain partners. That information is now stored in a way that makes it useable to different tools.

“We created a data lake, so we have access to all that data in a very efficient way,” says Papermaster. “This allows us to use AI in a multitude of fashions. We look at the data to find out where there’s a yield improvement based on interactions of our design with the manufacturing.”

For Papermaster, AI is proving to be very efficient to find needles in haystacks to improve chip yield and isolate bottlenecks in supply chain operations to determine where improvements can be made. “It’s really about the data and then creating the right team that can ask the right questions and put what they find out into action, improving manufacturing flow and increasing efficiency,” he says.

Another chip manufacturer, NVIDIA, agrees that semiconductors are an ideal application for AI. “Manufacturing chips requires over 1,000 steps, each of which needs to be performed to near perfection,” says Michael Kagan, the company’s CTO. “Sophisticated computations are performed at every stage to produce and pattern features the size of biomolecules. And AI is used to detect defects and monitor equipment.”

Other companies use the products they sell

Siemens delivers AI in their products and uses those same products for their own manufacturing processes in over 120 plants. They also deliver digital twins for customers to design products with less effort and material. That simulation software uses AI.

“As a player in the IoT field we offer software and connected hardware to connect the physical and the real world,” says Hanna Hennig, CIO of Siemens. “We use information technology and the injection of data analytics and AI to provide solutions, which not only automate the factory and the production line, but also actually make it adaptive—to become even more autonomous.

Hennig predicts that AI in manufacturing will only increase in the years to come. The machines and production lines will be able to change their configurations to manufacture different products, or different volumes. Sets of machines will be able to modify the way they produce things. “They may discover that, for instance, by altering the process carried out by a particular robot arm, they can minimize mistakes,” she says. “Or they might realize that rearranging a certain part of the production can help in cutting down waste.”

Siemens, in fact, recently partnered with NVIDIA to deliver AI-powered digital twin technology, combining the Siemens Xcelerator platform and NVIDIA Omniverse to allow customers to simulate processes and generate a massive amount of synthetic training data. The partners will also deliver tools to help process the synthetic data.

You say you want a revolution

While the companies experimenting with AI in manufacturing are enjoying some of the benefits, solutions currently require large doses of customization. CIOs who choose to jump on the bandwagon early pay for extra development costs and extra maintenance. “Given the inherent diversity of manufacturing assets, it would be a challenge for large companies to use a fully standardized approach,” says Cretella. “However, we’re standardizing platform components as a common denominator to enable specific use cases in each plant.”

For a very large company like P&G, it makes perfect sense to develop modules that can be reused for different product lines. But that approach wouldn’t be appropriate for smaller organizations, or companies with a narrower range of product lines. Most companies lack the resources to do a lot of customization and don’t have enough products to make it worthwhile to develop their own reusable components. Until more standard tools are available in the marketplace, the new era in manufacturing, with much lower production costs and far fewer errors, will have to wait.

In the meantime, the very large companies have a huge advantage. They can build their own platforms to use for different projects. “For AI to become a true differentiator, we needed to go beyond one-off initiatives and systematically scale algorithmic solutions across multiple company categories and in markets around the world,” adds Cretella. “With this program, we’re implementing the AI platform across hundreds of plants to enable specific use cases delivering quality, resiliency and optimizing water and energy consumption.” 

NVIDIA’s Kagan expects that when the industry does reach a level of standardization, it’ll revolve around models and frameworks. NVIDIA is already moving in this direction, offering some software packages to enable organizations to jump-start AI projects without highly skilled AI professionals.