Loading...

Digital Twins Become More Intelligent as AI Enters the Stack

Manufacturing IoT

Published by IOT San Jose Research & Editorial Team

Digital Twins Become More Intelligent as AI Enters the Stack

PepsiCo reports that a digital twin of its US manufacturing and warehouse sites, built with Siemens' Digital Twin Composer and Nvidia's Omniverse, helped it raise throughput by about 20 percent and lower capital spending by 10 to 15 percent. AI agents working inside the twin reportedly flagged up to 90 percent of potential issues before any hardware was changed, and design iteration cycles shrank from months to weeks. Those numbers are company-reported and early, and nobody outside PepsiCo has audited them. Even so, they describe a real change in what a digital twin is for.

The first generation of digital twins was a faithful mirror. It displayed what a machine or building was doing, sometimes in three dimensions, and a human decided what to do next. The new generation adds AI inside the twin, so it can propose changes, test them on a virtual copy, and learn from the result. The twin stops being a picture and becomes a workbench.

What AI Actually Adds

It helps to separate three jobs that AI performs inside a twin, because vendors often blur them together.

  • Making the twin. Building an accurate model used to take weeks of manual work. Tools such as Wandelbots' NOVA platform with Nvidia's neural reconstruction can turn scans and camera data of a real facility into a simulation-ready model much faster.
  • Running the twin. Physics-based simulation tells you what will happen. Machine learning can approximate that physics at far higher speed, so thousands of scenarios can be tested instead of dozens.
  • Reasoning over the twin. Language and agent models read the twin's data, answer questions in plain language, and recommend actions. ABB, for instance, links Omniverse libraries and Microsoft Azure into its Genix suite so operations teams can diagnose asset problems with AI agents.

Each job carries different risks. A bad twin is an accuracy problem. A fast approximate simulator can be wrong in ways that are hard to see. And a reasoning agent can state a confident conclusion that rests on a flawed model underneath. Buyers should ask which of the three a product actually provides.

What Announced in 2026 Tells Us

Several announcements at Hannover Messe in April 2026 show how the idea is spreading beyond factories. Kongsberg Digital added Omniverse to its Kognitwin platform for energy infrastructure, so operators can test scenarios before changing the physical site. Dassault Systemes uses virtual twins built on Nvidia physical AI libraries for what it calls autonomous, software-defined production. Microsoft combined Omniverse with Fabric Real-Time Intelligence for physically accurate simulations driven by live data. Humanoid reported that its HMND 01 robot completed autonomous logistics tasks in a Siemens electronics factory in Erlangen, and that simulation-first development cut hardware development time to seven months, against up to two years typically.

Notice the common thread. The twin is where the robot, the line, or the energy system is rehearsed. That is the same logic we described in our article on NVIDIA's physical AI push, where simulation and synthetic data reduce risky experimentation on real equipment. It also extends the ideas in our earlier piece on digital twins and predictive maintenance, where the twin was mostly a model of one asset.

GenerationWhat the twin doesWho decidesTypical data
MirrorDisplays live status in 2D or 3DHumanSensor readings, CAD
PredictorForecasts failures or performance from modelsHuman, informed by forecastHistory, sensor streams
SimulatorTests proposed changes in a physics modelHuman or software, with approvalPhysics models, live state
AgenticProposes, tests, and in limited cases executes changesAgent within limits, human oversightAll of the above plus operational context

Where the IoT Layer Matters

Every digital twin is only as good as the data feeding it, and that data comes from IoT devices. A twin of a distribution center needs door states, rack occupancy, and floor conditions. A twin of a chiller plant needs flow, pressure, and vibration. A twin of a building needs occupancy and air quality. When the sensor layer is incomplete or untrustworthy, the twin drifts away from reality, and an AI reasoning on a drifting twin becomes confidently wrong.

That is why the unglamorous work of sensor coverage, calibration, and naming conventions determines how far a twin can go. It is also why the most capable twin projects tend to begin with an audit of data quality, not with a visualization. For related reading on the data side, see our coverage of self-calibrating sensor fusion.

A Worked Example: Rehearsing a Layout Change

Consider a contract manufacturer in Fremont planning to add a new product line to a floor that already holds three lines. This is an illustrative scenario. The traditional approach is to draw the layout, argue over it in meetings, and learn what was wrong after the equipment is installed. With a twin, the team imports the facility scan, places virtual equipment, and runs the day's production flow with simulated operators and carts. The simulation shows a bottleneck at a shared inspection station and a cart route that crosses a pedestrian aisle.

An AI agent can then propose three alternatives, run each across a range of demand levels, and rank them by throughput and safety clearance. The team picks one, and engineers check the reasoning before approving. The value is not that the agent is smarter than the engineers. It is that the agent explores many more variations than anyone has time to draw by hand, and it does so before concrete is poured or conveyors are bolted down.

The same approach applies to construction projects, a theme explored in our article on construction digital twins in San Jose, where rehearsing a build sequence can expose clashes between trades before crews arrive.

Questions to Ask a Twin Vendor

Demonstrations of digital twins are persuasive because they are visual. Ask questions that a good picture cannot answer. How is the twin kept in sync with the real facility, and how often? What accuracy has been measured against reality, and by whom? Which parts of the model are physics-based and which are learned approximations? Can the data and models be exported in open formats? What happens to the twin when a machine is replaced or a sensor fails? And which customer results can be shared with enough detail to understand what was measured? Vendors confident in their product will answer these directly. Vague answers suggest a presentation layer without much underneath.

It also helps to ask about people. A twin needs an owner who updates models after layout changes, reviews accuracy, and decides what the agents may do. Projects without that role tend to decay within a year, as the real plant moves on and the virtual one stays behind.

A Note on Where Twins Fall Short

A twin is a model, and every model leaves something out. Friction changes with wear, operators improvise, and suppliers deliver parts that differ slightly from the specification. A simulation that ignores those details can rank two layouts correctly most of the time and fail badly in the one case that matters. The remedy is to validate the twin on a past event whose outcome is known, and to treat disagreements as information about the model, not as noise. Teams that do this build credibility, and credibility is what allows an agent's recommendations to be acted on.

Business Value

  • Lower capital risk. Finding layout and capacity problems virtually costs far less than finding them on site.
  • Faster iteration. PepsiCo's reported shift from months to weeks for design cycles shows the direction, even if the exact figure will vary by project.
  • Safer robot deployment. Testing robot behavior in simulation reduces exposure of people and equipment to untested software.
  • Better knowledge capture. A twin that records decisions and results preserves expertise that otherwise leaves with experienced staff.

Implementation Advice

  1. Pick one decision to improve. Define a specific choice, such as where to place a new line or when to service a compressor, and build the twin to support it.
  2. Audit the data first. Check sensor coverage, timestamps, units, and naming before investing in 3D visuals.
  3. Measure the twin's accuracy. Compare predictions with reality regularly, and recalibrate when they diverge.
  4. Keep humans in approval roles. Let agents propose and test, and require sign-off for real-world changes until trust is earned.
  5. Avoid lock-in where you can. Prefer open data formats and interfaces so the twin survives a change of vendor.

Common Challenges

ChallengeWhy it appears
Model driftThe real facility changes while the twin stays still, so predictions degrade
Unverified resultsMany published gains are company-reported and measured on selected projects
Cost of high-fidelity modelsDetailed twins demand modeling effort and computing resources
Data integrationPlant systems from different eras and vendors rarely share a common data model
Over-trust in agentsA fluent explanation from an AI can hide a flawed assumption in the model

What Comes Next

Expect twins to become less of a special project and more of a standard layer in design, operations, and training, with simulation, live data, and AI agents sharing one environment. Expect the debate to move from whether twins work to how their claims can be verified. For readers in the South Bay, where manufacturers, robotics developers, and facility operators all live within a short drive, the practical step is modest: choose one high-value decision, build a small twin around it, and measure whether it improves the outcome. For the compute that powers these twins, see our article on data center growth and monitoring.

Frequently asked questions

Earlier twins mainly mirrored live status for humans to interpret. AI now helps build twins faster, run simulations at much higher speed, and reason over twin data to propose and test changes, turning the twin from a display into a workbench.

PepsiCo reported about a 20 percent throughput increase and 10 to 15 percent lower capital spending using a Siemens Digital Twin Composer environment built with Nvidia Omniverse. These are company-reported early results that have not been independently verified.

It is a Siemens platform that combines Xcelerator digital twin data with Nvidia Omniverse to create real-time, photorealistic virtual environments where design, simulation, and operations can be tested together before physical changes.

A twin depends on sensor data to stay aligned with reality. Missing coverage, bad calibration, or inconsistent naming makes the twin drift, and an AI agent reasoning on a drifting twin can reach confident but wrong conclusions.

Pick one specific decision to improve, audit sensor data first, measure the twin's accuracy against reality, keep humans approving real-world changes, and favor open data formats to limit vendor lock-in.

Model drift, unverified vendor-reported results, high modeling cost, difficult data integration, and over-trust in AI explanations that may rest on a flawed model. Human sign-off and regular accuracy checks reduce these risks.

Related articles

Top