Digital Twins

A digital twin is a virtual model of one specific physical system that stays connected to that system through live data. As the physical asset operates, sensors feed the model state information, and the model updates to match. In its fuller form, the connection runs both ways: something learned in the virtual model, a predicted failure, an optimized setting, gets pushed back to change how the physical system is run (Glaessgen & Stargel, 2012; International Organization for Standardization, 2023).

The idea did not start as a buzzword. NASA leaned on detailed vehicle simulators to diagnose the Apollo 13 oxygen tank failure and test fixes before relaying instructions to the crew, an early instance of using a virtual stand-in to solve a problem on the real system in real time (NASA, 2025). Michael Grieves gave the concept a name and a formal structure in a 2003 University of Michigan product lifecycle management course, and NASA's John Vickers is generally credited with coining the term "digital twin" itself around 2010 (Grieves & Vickers, 2017; NASA, 2025). The concept went mainstream after Grieves published a 2014 white paper framing digital twins as a manufacturing tool, around the same period NASA and the U.S. Air Force were formalizing a version of it for aerospace vehicle sustainment (Glaessgen & Stargel, 2012; Grieves, 2014).

Systems engineers already have a map for where value gets created, and lost, across a program: the Vee model, decomposition down the left leg, integration and verification climbing back up the right. Digital twins don't sit at one point on that V. They attach at three, and each attachment does a different job.

None of the three is optional if the goal is a program that learns from its own operational history instead of repeating the same design mistakes on the next iteration.

A digital twin's entire value depends on one assumption: that the model still represents the real system. That assumption erodes constantly, and quietly. A field technician swaps in a part with a slightly different tolerance. A firmware patch changes how a controller behaves without anyone updating the matching parameter in the model. A component wears at a rate the original model never accounted for. None of this throws an error. The twin keeps producing a dashboard and a set of predictions, and it just becomes progressively wrong (Mertens & Denil, 2025).

Mertens and Denil (2025) make the point plainly in their case study of a twinned gantry crane: physical systems are not immutable once deployed the way software is. They evolve through maintenance, wear, and user error, so a twin that isn't continuously checked against fresh telemetry will keep drifting without anyone noticing, until the gap is large enough to cause a bad decision. Their fix borrows directly from model-based design: reuse the same validation metrics used to trust a simulation model in the first place, run them continuously instead of once at handover, and flag the twin for recalibration once it drifts past a set threshold.

This is the same discipline the Vee model already formalizes for the rest of the program. Verification and validation don't stop being relevant once a twin exists. They just move from a one-time gate at delivery to a running check against the twin itself, for as long as the twin stays in service.

A twin that has silently drifted out of sync and is still being trusted is a worse position to be in than having no twin at all, because the dashboard still looks authoritative.

The value case splits into two buckets: money saved by testing virtually instead of physically, and money saved by catching problems before they become failures.

Faster, cheaper iteration. Companies building product digital twins report up to 25 percent fewer quality issues once a product reaches production, and one company reported 3 to 5 percent higher sales on digital-twin-based products thanks to better features and quality (McKinsey & Company, 2023). At the factory level, one manufacturer used a factory digital twin to redesign its production schedule and cut monthly costs by 5 to 7 percent, while also surfacing hidden bottlenecks the schedule change hadn't been built to find (McKinsey & Company, 2024a).

Aftermarket and service revenue. Once a product twin exists from the design stage, it can support predictive maintenance and in-service optimization as paid services, lifting revenue by 5 to 10 percent in some product categories (McKinsey & Company, 2023).

Materials and sustainability. Consumer electronics manufacturers using digital twins to catch design and process waste earlier have cut scrap by roughly 20 percent (McKinsey & Company, 2024b).

Adoption as its own signal. Seventy percent of C-suite technology executives at large enterprises are already exploring or investing in digital twins, according to McKinsey's own research, which says as much about expected value as any single case study does (McKinsey & Company, 2024b).

Standardization is closing a decade-long gap. The field spent years with every vendor defining "digital twin" its own way. ISO/IEC 30173:2023 now gives the field shared terminology, ISO 23247 already standardizes the manufacturing use case, and INCOSE's October 2024 liaison with the Digital Twin Consortium is an explicit attempt to fold digital twin practice into recognized systems engineering methodology instead of leaving it as a parallel discipline (Digital Twin Consortium, 2024; International Organization for Standardization, 2021, 2023).

Twins are starting to compose. Individual asset twins are combining into larger structures: a fleet twin built from many vehicle twins, a plant twin built from many machine twins, a mission twin built from platform twins. Keeping every layer consistent when the pieces update at different rates and different fidelities is now its own active research problem rather than a solved one, and it inherits the same drift problem described on Sheet 3, at a larger scale.

The concept is spreading past hardware. Gartner's 2025 Hype Cycle work places the industrial digital twin well past its early hype and into practical deployment, while newer variants like a "digital twin of a customer," a model used to simulate how a person might respond to a given offer, are still sitting at the earliest, least proven stage of the cycle (Gartner, Inc., 2025). Watch for the same twin-and-sync logic that started on spacecraft and factory floors getting tried on customers, employees, and cities. Not all of those attempts will hold up.

Physical AI convergence. Chip and simulation vendors are pushing digital twins toward becoming the training and proving ground for embodied AI and robotics, not just a monitoring dashboard. Recent industry partnerships pair physics-based simulation, digital twins, and agentic AI to build and test what they call physical AI systems before any hardware exists (RTInsights, 2026).

A digital twin without AI is a very good mirror. Someone still has to look at the dashboard, notice the trend, and decide what to do about it. Layer AI on top and the twin stops being purely reflective and starts doing part of that interpretation itself (Grid Dynamics, 2026).

Three developments are doing most of the practical work right now. Generative AI is cutting the time it takes to build a twin in the first place, turning sparse scans or design data into usable 3D models and parameter sets (Gebreab et al., 2024). Large language model agents are being used to parametrize and adjust simulation models faster than an engineer scripting the same change by hand (Xia et al., 2024). And in domains that don't have enough real-world data to train a purely data-driven model, physics-informed neural networks fold known engineering equations directly into the AI model, which keeps predictions physically plausible even on sparse data (Liu & David, 2025).

What this doesn't fix. None of this removes engineering judgment from the loop, and it shouldn't. Three problems show up in nearly every serious deployment. Explainability: a black-box model recommending an action is hard for an engineer or a regulator to sign off on if it can't say why (Grid Dynamics, 2026). Integration cost: connecting an AI-enabled twin to legacy ERP (enterprise resource planning), MES (manufacturing execution systems), and OT (operational technology) systems is usually harder and more expensive than the AI model itself. And compute cost: running continuous simulation, generative models, and agents at the same time is not free, and it shifts budget toward infrastructure and away from other priorities. Even vendors pushing hard on autonomous, "agentic" engineering workflows describe the current reality as a layered process, generation paired with a verification step, rather than a fully autonomous decision-maker, and are explicit that responsibility for the outcome still sits with a human engineer (IndexBox, 2026).