What happens when AI redesigns a car using information that is already out of date? Supply chains can break overnight, software is constantly updated, and vehicle configurations and production conditions keep changing. In this environment, knowing what is true right now matters as much as the power of the AI itself. Siemens believes the answer lies not in building a bigger AI model, but in data that continuously tracks every change to a product across design, simulation, production, and operation. We followed that message from the wildfire that opened Realize LIVE to the factory floor.

By Sang Min Han  _ han@autoelectronics.co.kr
한글로보기













Wildfires and the Present Tense

Detroit, June 1. Bengaluru, August 4. On both stages of Realize LIVE, Tony Hemmelgarn, President and CEO of Siemens Digital Industries Software, opened his keynote with the same story. A wildfire.
That wasn't a coincidence. For Hemmelgarn, the wildfire was more than a dramatic opening. It captured the question Siemens wanted to put to its customers worldwide. Follow why he reached for the same image across two continents, and it becomes clear what Siemens wants AI to actually do.
A wildfire map doesn't survive the day. Satellite imagery, fuel maps, wind direction, crew positions, spread-prediction models — all of it updates by the minute, and the safe zone drawn yesterday turns into a trap the moment the wind shifts.
"Yesterday's data is extremely dangerous."
The problem isn't a shortage of information. There's plenty of it. The problem is whether you can pull it into one picture of right now.
Hemmelgarn then applied the same logic to manufacturing. Supply chains break overnight. A material available yesterday can vanish tomorrow. Production sometimes has to move from one region to another with no warning. The product itself isn't fixed either — machines that used to be purely mechanical are now systems built from semiconductors and software, and their configuration keeps changing after they ship.
So what industry needs isn't an unshakeable, perfect plan. It's the ability to change the plan the instant reality changes. A firefighting crew surviving a wildfire and a manufacturer surviving a supply-chain collapse are facing the same question: what just changed, and what does it mean for where I'm standing right now.
Where, then, should AI look for industry's "now"? Siemens' answer isn't a new AI model. It's the more than 20 years the company has spent building PLM.







Where AI Finds Something to Trust

Siemens' AI strategy starts with Teamcenter.
"AI only really works when it's built on Engineering Truth. Teamcenter is where customers go to get data that's accurate, trustworthy, managed, and connected."
Hemmelgarn named Teamcenter as the backbone for managing Engineering Truth.
There's a practical reason for that. AI runs on data, but on the factory floor, having a lot of data doesn't mean much by itself. You need to know which product version it belongs to, why it was designed that way, which parts it connects to, whether the values have been validated against real physics and simulation, who's allowed to see those attributes, and whether an AI's judgment can actually flow into an engineering change and production. A single data lake rarely checks all those boxes.
Joe Bohman, executive vice president of PLM products, gave an example from conversations with customers. CIOs tell him they've settled on platforms like Snowflake, Databricks, or Palantir. Talk to the engineering teams at the same companies, though, and the story looks different — they're struggling to get real use out of those tools, because engineering data is too complicated for them.
Hemmelgarn made a similar point his own way, comparing product data to a runner. "Take a snapshot and you'll get oxygen consumption, pace, distance. What you won't get is what happens two minutes from now." Most data lakes, he said, never get past that snapshot. His conclusion: trustworthy results need context, not just data.
Siemens is pitching its new Intelligence Center X on two points that follow from this. Through Teamcenter, it can trace the latest product configuration and the context behind every change. And that configuration stays alive instead of freezing into a snapshot. This is where Siemens draws the line between itself and general-purpose data platforms in industrial AI — engineering data, and the context wrapped around it.
In a one-on-one interview on the first evening of the event, I asked Sam Mahalingam, executive vice president of simulation, HPC, and AI, whether the day's message really came down to Engineering Truth. "Yes, that's exactly right," he said.
"Palantir can add great context to business data," he added, "but the engineering world has to be grounded in Engineering Truth."
Hemmelgarn said interest in and growth around Teamcenter has clearly picked up over the past 18 months. As companies move from talking about AI to actually deploying it, he said, many are discovering — often later than they'd like — that they never had a reliable data foundation to begin with. Rather than making PLM less relevant, generative AI has brought it back to the center of industrial strategy.



A Product That Stays Alive, Not a Snapshot

At Siemens, a product is no longer a single machine.
Ankur Gupta, head of integrated circuits, opened with a number: today's AI accelerator chips pack roughly 200 billion transistors onto a single die. Stack semiconductors at that density in three dimensions and heat gets trapped inside. How that heat moves, as one continuous physical process, was the center of Gupta's example.
Siemens' EDA tools build thermal models of the chip and package down to the nanometer. That model then feeds into Simcenter, which works out thermal behavior at the board and system level. The results loop back into the heat-sink and enclosure design, and once the product is running, operations software ties into sensor data to control cooling. Siemens calls this chip-to-system: one physical phenomenon moving without a break from chip to product to operations.
That connection doesn't stop at design and simulation. In Bengaluru, Gupta described a real data center where chips lower their supply voltage — and their power draw — based on the conditions and workload they're actually facing. Chips are normally designed with margin built in for worst-case conditions, but real operating environments rarely stay at that extreme. Read the actual state through sensors and software, and that margin becomes something you can spend on power savings instead. Gupta said the approach measured 14 percent in power savings at an actual data center.
The same approach in a car, he said, could improve battery efficiency and extend driving range. On August 11 in Seoul, I asked Gupta whether this chip-to-system optimization could also cut power consumption in real time inside an actual vehicle. "It can," he said, explaining that the method lowers the supply voltage to match the chip's actual condition, within a range that doesn't trigger functional errors. He added that chips are designed around worst-case conditions that don't always match real-world use, and that Siemens is working with proteanTecs, which brings the sensor and software technology for this kind of monitoring.
Designing a single semiconductor, it turns out, now means understanding the entire system that chip will live inside.
That is why PLM has to take on EDA, software requirements management (ALM), electrical design, and simulation all at once. What Siemens calls the comprehensive digital twin means connecting semiconductors, PCBs, wiring harnesses, electrical and electronic architecture, mechanical geometry, software, simulation, manufacturing, and operations on a single dataset. Showing off one chip or one mechanical shape in isolation is not what Siemens means by that term.







The Layer That Turns Data Into Judgment and Action

Managing all that data is one problem. Acting on it is another. Intelligence Center X, which Siemens launched two months ago, is meant to close that gap. Bohman described it this way:
"Intelligence Center X is a single, self-contained agentic development environment built to understand these AI-native industrial applications."
The architecture consists of three layers. At the bottom sits an ontology and knowledge graph that connects the meaning and relationships of data scattered across Teamcenter, ERP, CRM, and MES. Above that sit industrial machine learning, physics-based AI, and large language models. At the top, people and AI agents carry out the actual work, including engineering changes.







From Insight to Engineering Change

How this data actually moves showed up in an example Hemmelgarn gave during his opening keynote. A global aircraft manufacturer kept seeing unplanned flight interruptions across its jet fleet, traced to hydraulic leaks and pressure spikes when the landing gear deployed. At first it looked like a problem with individual aircraft. Once the comprehensive digital twin — aerodynamics, structures, and hydraulics analyzed together — was brought in, a different picture emerged: turbulence was disturbing the load on the landing-gear door, and that variation was what triggered the pressure spikes downstream. Not a run of isolated failures. One recurring problem.
Hemmelgarn used this example to explain the role of Intelligence Center X.
"We combine that fleet's flight data with physics-based simulation to pinpoint the root cause." Once Teamcenter identifies every aircraft configuration exposed to the same issue, Intelligence Center X weighs the options: redesign the part, adjust the maintenance interval, or accept the cost. Insight becomes a decision, and the decision becomes action — an agentic workflow that carries it through end to end.
The next day, Bohman picked up the same example and ran the workflow live.
He asked Teamcenter Copilot to "help release this change," and impact analysis started immediately. Work that would normally take hours, sometimes days, finished in seconds. From there he moved into Intelligence Center X, which pulled alternative parts from across the company's data to lower supply-chain and supplier risk. Finally, it located and linked the simulation data needed to verify whether the hydraulic-system change would actually hold up, confirming it was safe to carry into production. Most analytics platforms stop once they've delivered insight. Siemens is trying to build the loop back — from insight into engineering change, and from there into the production plan.








Physics AI: Narrow It Down Fast, Verify It With Physics

Physics AI doesn't replace computation. It picks where to spend it more intelligently.
"The physics engine itself doesn't change. What changes is how smart you are about pointing it. This isn't about replacing deterministic truth — it's about making that truth scale."
Sam Mahalingam, who leads simulation, HPC, and AI, gave the example of Hyundai Motor. Optimizing a truck-frame subsystem — cab, deck, and suspension all attached — used to take a week. With Physics AI, it took 15 minutes. Feed in the CAD geometry and it returns a bending-displacement prediction on the spot; checked against real CAE results, the accuracy landed around R² 0.98 to 0.99.
I asked whether that kind of speed could be trusted in automotive design, where safety and regulation are on the line. Mahalingam drew a clear boundary: Physics AI doesn't have the final word.
"Physics AI helps you explore a complex design and sweep through a huge design space very fast. Once you've narrowed it to one or two, or three or four candidates, that's when you bring in a physics-based solver for real verification. Physics-based simulation will still be the final check for certification."
He noted it took nearly 20 years for finite element analysis to become an accepted certification tool in the auto industry, and Physics AI won't become one anytime soon. But once enough validation data builds up, he said, Physics AI could take on some role in certification within five or six years.







Where the Loop Closes: The Factory

Siemens' electronics plant in Germany shows how this loop of trust closes on the ground.
Volker Albrecht, who leads Siemens Digital Logistics, said the company treats its own plant as a proving ground for industrial AI, and walked through a project built with Nvidia. When an anomaly showed up on a surface-mount (SMT) line, a floor worker asked an agent what was causing it.
"There are three possibilities: the machine, the process, or the design." The agent answered, asked whether to keep digging, and worked through all three. The cause turned out to be a design flaw in an RJ45 connector — the surface-mount version needed to switch to through-hole mounting because of a mechanical-strength issue. The agent then asked whether to open an improvement workflow in Teamcenter. Once approved, the actual engineering-change process began.
That wasn't the end of it. Change one design, and you have to check what that does to the whole line. Running the plant's full line through the Digital Twin Composer, built with Nvidia, turned up a new bottleneck in material flow. That simulation can run hundreds of thousands of times, and the results go straight into production decisions. Albrecht said the plant is expected to lift productivity by 7 percent every year, and repeated simulation like this is what makes that target reachable.

Set this next to the wildfire and the same pattern emerges. On a fire: detect that the wind has shifted, pull the current situation into a single picture, judge the impact, change the plan. On the floor: detect an anomaly, pull up the latest product configuration and design intent stored in Teamcenter, judge the cause, change the design, then check the result through simulation again.
The difference isn't whether AI produces an answer. What matters is whether that answer rests on the current product configuration and design intent, on data that's been physically validated — and whether it can turn into an actual engineering change and production action. Siemens' version of industrial AI doesn't end at generating insight. It verifies that insight, carries it into action, and feeds the result back into the data.

Yesterday's data can't run today's factory. That's why Siemens keeps PLM at the center, even in the age of AI.



From left: Tony Hemmelgarn, President and CEO; Joe Bohman, EVP of PLM products; Sam Mahalingam, EVP of Simulation, HPC, and AI; and Ankur Gupta, head of Integrated Circuits — all of Siemens Digital Industries Software, photographed at Realize LIVE Asia-Pacific.

AEM(오토모티브일렉트로닉스매거진)