Digital Twins and Their AI Use Cases

Barry Boehm, perhaps the most influential software engineer of his generation, liked to tell the story of a weights engineer who was responsible for tracking the weight of everything on a spacecraft. The weights engineer noticed a multi-million-dollar budget item for software, so he asked a software engineer how much the software weighs. The software engineer replied, “Nothing.” Weeks later, the software engineer brought back a big deck of punch cards. The weights engineer told the software engineer that he had to weigh the deck before it went on the spacecraft, to which the software engineer replied, “We only use the holes.”[1]

Forty years later, Google began training cars to drive themselves. They built a fake city in the desert for the endeavor but  it took too long to reconfigure. So they built a virtual simulator in which their cars’ software learned faster than in real time, and Google engineers created hundreds of variations to each configuration, a practice they unpretentiously called “fuzzing.”[2]

Boehm’s punch card holes and Google’s fuzzing are examples of the long-term trend of digitization. Digital twins are the application of digitization to industrial processes and are generating massive ROI in many industries. 

A digital twin is a virtual, digital representation of a physical object or process that we can use AI on to rapidly and cheaply get answers about the real-world object or process. Most companies obtaining high ROI from digital twins use a crawl-walk-run approach since it often requires cultural and mindset changes. Here are some use-case examples:

Crawl

A 130-year-old food sciences company, Bunge Loders Croklaan, wanted to accelerate their design process to keep up with changing consumer behavior caused by COVID-19. The company specializes in plant-based oils and fats and has prototyped thousands of products over its history. Their goal was to leverage data from those past prototypes to quickly predict attributes like color and viscosity of new designs. They used AI and deployed their model in just two weeks. “Today’s tools are amazing. Lightning speed results. They enable play, enable fun,” said , Ph.D., Bunge’s Director of Innovation. [3]

Oil fields are often in remote, inhospitable locations. It can take a long time for repair engineers and parts to arrive when something breaks. In order to maximize uptime, one operator developed a predictive maintenance model for oil well power drives. Some of their field engineers expected cutting and kinetic parts to be the highest maintenance but that wasn’t the case at all. Their AI model showed that circuit boards, operating in hot, dirty environments, are the main cause of unplanned downtime. They expect the model will reduce equipment failures by 85%.  

Walk

GE Aviation wanted to accelerate the design of new jet engine parts to make them more sustainable, namely, reduce soot emissions and increase fuel efficiency. Two generations ago, they would fabricate a new part, put it in an engine in the shop, run it for a while, and measure efficiency and emissions. The process took weeks. Computing costs dropped and they switched to computational fluid dynamics, a digital simulation of physical properties, which used expensive computers and took days. Computing costs plummeted so they developed deep learning AI models to predict fuel efficiency and soot emissions of digital designs. It’s 190 million times faster than computational fluid dynamics and allows them to test thousands of tweaks to a design in minutes. [4]

Pharmaceutical companies are using AI to detect diseases years earlier. Research indicates that AI can detect Alzheimer’s disease in PET scan images 6 years earlier than today’s methods. Merck is working on using mobile phone accelerometer data to detect the shuffling gait and small sets characteristic of Parkinson’s disease.

Run

NXP, a multinational $8.6B/year semiconductor manufacturer, uses AI to detect when manufacturing line yields are dropping, in near real time. Previously they pulled a product off the line every three days and rigorously tested it. It could take up to three days to detect a problem and all products produced since the problem occurred were defective. With AI they monitor thousands of parameters of the manufacturing process and every few minutes predict whether the process is generating defective products. This drastically reduced the time to detect defects and thus the number of defects. [5]

The last example is also from semiconductor manufacturing and has maybe the highest AI ROI I’ve ever seen outside of financial trading. Manufacturing processes for high-end chips are complex, with tens of thousands of steps and thousands of parameters to be tuned. The machines used generate terabytes of data each day. The design of the manufacturing process for a new chip is simply called “the recipe.” Key steps to getting to market are designing and testing the recipe in an R&D lab, and installing and testing the recipe to fabrication facilities (so called foundries) around the world. Foundries are offline during this time and not making money so reducing the duration of these steps is valuable. One semiconductor equipment manufacturer applied AI to predict the effect of recipe changes without testing them in physical machines, cut three to six months off the process, and saved $1B per foundry per chip design! And they’re in a lot of foundries. 

Getting started

Simulating the real world may seem daunting, but cloud computing and today’s AI algorithms are well up to it. Best practices for starting a digital twins initiative are similar to those for any AI initiative: [6]

• Set clear goals and business ROI expectations

• Get a champion in the executive suite

• Avoid moonshots for your first project

• Get quick wins, document them, and evangelize them widely

• Use interdisciplinary teams of subject matter experts, data scientists, and business analysts

• Collect data from every possible source, assign data owners, and hold owners responsible for quality, velocity, and other SLAs

[1] "ACM Fellow Profile Barry Boehm," Association of Computing Machinery, Software Engineering Notes, vol. 29 no. 5, September 2004 

[2] “Inside Waymo's secret world for training self-driving cars,” Alexis Madrigal, The Atlantic, August 23, 2017

[3]  “Digital Innovation in the food manufacturing industry,” BrightTALK, Sept. 10, 2021

[4] “GE says it’s leveraging artificial intelligence to cut product design times in half,” Jeremy Bogaisky, Forbes, March 6, 2019

[5] NXP Semiconductors - Reducing Detection Time of Manufacturing Issues with Real-time Automated Process Control

[6] For more best practices see the Dataiku ebook, The Rise of Industry 4.0: Drive Efficiency With the Power of AI and the IoT