Ask one question of anything sold to you as a digital twin. When the physical thing changes, does the model find out on its own — and when the model concludes something, can it act back on the physical thing? Two yeses is a twin. One yes is a shadow, and zero is a drawing with a subscription fee.
That is not a heuristic I invented; it is close to the field’s own formal test. In a much-cited 2018 classification, Kritzinger and colleagues sorted the “digital twin” literature by the direction and automation of its data flows — and found that much of what is published, let alone sold, under the name fails the definition:
So: a digital twin is a virtual representation of one specific physical thing — an engine, a factory, a patient, a city — bound to it by a live data link, so the virtual copy tracks the physical one’s actual state and can feed decisions back. The link is the twin — not the model, not the 3D render — and that one clarification sorts most of the market’s claims into honest and otherwise.
The concept has a cleaner paper trail than most buzzwords. Michael Grieves drew the essential structure — a real space, a virtual space, and data flowing between them — on a product-lifecycle-management slide in 2002, and spent years calling it things nobody remembers (“Mirrored Spaces Model”). The name that stuck came from NASA’s John Vickers, and the definitive formulation arrived in a 2012 NASA/Air Force paper by Glaessgen and Stargel: an integrated, multiphysics, probabilistic simulation of an as-built vehicle, updated by sensor data and fleet history, that “mirrors the life” of its flying twin. Note the words as-built and probabilistic — not the design, but this serial number, with its dents; not one prediction, but a distribution. The 2012 definition was already more honest than most 2026 marketing.
And yes, about Apollo 13. The story that NASA’s ground simulators — hastily reconfigured to match the crippled spacecraft — constituted “the first digital twin” is told in practically every vendor deck, mine-adjacent industries included. The simulators were real and the improvisation heroic; the framing is retrospective, applied by digital-twin marketers decades later, and I could find no primary NASA source using any such concept at the time. It is a good story about simulation. It is not evidence about products.
That definition also set a bar the technology still cannot clear. The paper that says so most plainly — and the one I would make required reading for anyone buying a twin — is Rasheed, San and Kvamsdal’s “Digital Twin: Values, Challenges and Enablers”: high-fidelity physics simulators are orders of magnitude too slow to run in real time against a live asset. The twin’s defining feature — liveness — is exactly what the best models can’t afford.
Which is why a twin is less a static artifact than a live promise: that the virtual state tracks the physical state closely enough to act on. Physical things drift. Components wear, sensors decay, operators improvise, last month’s calibration quietly stops holding. An unmaintained twin doesn’t fail loudly; it becomes a confident description of an object that no longer exists — and everything decided on it inherits the gap.
This is where the stochastic machinery from earlier in this series re-enters. A serious twin is probabilistic by construction — NASA’s 2012 definition said so — which means state estimation under noise, ensembles rather than single runs, and surrogate models (increasingly neural ones) standing in for physics that can’t run at wall-clock speed. The research frontier is precisely this trade: how much fidelity can be compressed into a model fast enough to keep the promise of liveness.
Sorted by source type, the benefits literature splits cleanly, and almost none of it is what it looks like.
Named-company results are reported by the companies. Rolls-Royce executives credit engine-health twins with substantial fuel and CO2 savings — including, per trade-press interviews, a 48% increase in one engine’s time-on-wing between maintenance removals. Siemens publishes customer cases with hard numbers. These are real engineering programs, and self-reported figures, unaudited by anyone.
The most-quoted number in the field is sponsored content. GE’s famous claim that digital wind farms boost annual energy production “up to 20%” traces to a 2015 sponsored placement in Harvard Business Review — literally marked “Sponsor Content” — that a decade of citations has quietly laundered into a research finding. Nothing I found refutes it; nothing independent confirms it. It is an advertisement with excellent citation metrics.
The market-size numbers deserve open scorn. Estimates for the 2025 digital-twin market alone span roughly $17–40 billion depending on the research mill, with 2035 forecasts diverging by five to ten times. An order-of-magnitude spread means invention, not measurement. I will not be citing one.
What survives the labeling is the peer-reviewed critical literature: systematic reviews documenting the field’s definitional sprawl, and the liveness limit described above. The hard parts get published; the savings get announced.
Three frontier threads survive the same scrutiny. Industrial twins at building-scale are real and commercially deployed — NVIDIA’s Omniverse customer roster (BMW’s factory twins, Schaeffler’s robot-planning simulations) is vendor-published but names real programs at real companies. Medicine is placing the boldest bet: the European Commission’s Virtual Human Twin initiative — €100 million, launched December 2023, ninety-plus signatory organizations — aims at patient-specific physiological twins for testing treatments in silico before touching the patient. And cities have quietly run the longest experiment: Singapore’s national 3D twin has been operational since the mid-2010s, now used for flood and urban-heat modeling. The newest thread — agentic twins, where the twin doesn’t just mirror but decides — is where this series’ two threads converge: a live model of the world, plus an agent licensed to act on it, plus a human owning the consequences.
So: what is a digital twin, really? A maintained promise of synchronization between a probabilistic model and one specific piece of reality — expensive to keep, easy to fake with a render, and, when actually kept, the closest thing the physical world has to the context engine I keep arguing every AI system needs. The buzzword will pass. The discipline of keeping models honest against reality was valuable before the name existed — ask NASA, 2012 — and will outlive it.
References