Two federal documents published in 2024 define a digital twin, and the definitions are not compatible. Reading them side by side is the fastest way to understand why this subject generates so much noise and so few evaluations.
The National Academies committee defined a digital twin as a set of virtual information constructs mimicking the structure, context and behavior of a physical system, dynamically updated with data from that system, carrying a predictive capability, and informing decisions that realize value. Its Conclusion 2-1 reduces the definition to two requirements: modelling and simulation producing a virtual representation of a physical counterpart, and a bidirectional interaction between the virtual and the physical.
FHWA’s Resource Center, in a fact sheet on digital as-builts issued the same year, offers this: a digital twin is a digital representation of an asset that is geospatially located within the system and connected to relevant inventory, condition and performance information, providing what the fact sheet calls a “one-stop-shop” for the history, status and performance of each asset. The verbs in the surrounding passage are conditional throughout. Digital twins would provide a single reference. As-designed information could be stored and then verified during construction. The graphic beside the text is labelled as the Utah Department of Transportation’s vision.
There is no bidirectional loop in the highway definition and no predictive capability. By the National Academies’ test, what the highway sector calls a digital twin is a well-organised asset database with a map attached, a useful thing to build but not the thing the word was coined to describe. Definitional drift of that kind costs the technology section more than any single overstated pilot does, because it makes the pilots impossible to sort.
The committee’s own warning about the evidence
The National Academies report does not treat the terminology drift as harmless. Its assessment is that “the publicity around digital twins and digital twin solutions currently outweighs the evidence base of success,” and that it is challenging to separate what is true from what is merely aspirational, because there is no agreement across domains about what the term covers.
The technical version of the same concern is sharper. Verification, validation and uncertainty quantification are the disciplines that establish whether a model can be trusted for a decision, and the committee found a gap between the class of problems addressed in conventional modelling and simulation and the problems that arise when a model is continuously updated from field data and used to inform action. A model that changes every time new sensor data arrives is a model whose validation has to be continuous too, and nobody has settled how to do that.
For infrastructure that matters more than it does in most fields, because the decisions on offer are load posting, inspection interval and repair sequence, and each of those has a public safety consequence and a legal record behind it.
What highway agencies have actually deployed
The concrete program is digital project delivery, and FHWA promotes it through the eighth round of its Every Day Counts initiative rather than as a research effort. The named adopters are specific. Arizona, Oklahoma and Utah have developed multi-year digital delivery plans. Illinois and Arizona rewrote survey and design manuals to accommodate modern workflows. Pennsylvania and Iowa are testing three-dimensional model-based project delivery on complex bridge projects.
FHWA’s account of the barriers is more informative than its account of the benefits. The problems it names are paper and PDF workflows that force manual re-entry of data, siloed databases, incompatible software, and manuals that predate lidar, three-dimensional modelling and uncrewed aircraft. All of it is a records and file-format problem rather than a modelling one, and the digital as-builts fact sheet is explicit about where the constraint bites: three-dimensional model information is searchable and extractable only when opened in compatible proprietary software, which is why AASHTO has resolved to adopt Industry Foundation Classes, the open standard published as ISO 16739, with version 4.3 as the target that would make the models readable in any conforming product.
An agency that cannot read its own model in twenty years does not have an asset record. It has a license. That constraint is upstream of every twin ambition in the sector, and it is the reason the serious work in this field looks like standards adoption rather than simulation. How the model is produced in the first place depends on the contract, which is covered in project delivery methods.
FHWA’s test for a digital as-built reads as the minimum specification for anything grander. The record has to be electronic, so it can sit on a network or be sent. It has to be searchable, so information can be found inside the file. It has to be extractable, so a system can pull data out without a human retyping it. And it has to be durable, meaning accessible over the life of the asset, which for a bridge is longer than the commercial life of most software.
Ranked against those four, the tactical options the fact sheet lists sort themselves cleanly. Vector mark-ups on PDF plan sets conform to ISO 32000-1:2008 and keep their text and geometry extractable, where a raster mark-up is only a picture. Field survey points and geographic information system datasets are searchable, extractable and stored in durable database formats. Three-dimensional models score well on the first three properties and fail the fourth until the open standard displaces the proprietary reader. An agency that wants a twin later is, in practice, choosing file formats now.
The instrumented bridge that predates the vocabulary
The replacement for the collapsed I-35W bridge in Minneapolis was instrumented from the start. When the St Anthony Falls Bridge was rebuilt, the design team installed a monitoring system that combined vibrating wire strain gauges, thermistors, linear potentiometers, accelerometers, concrete corrosion sensors, humidity sensors and long-gauge fiber optic deformation sensors. The 2009 paper describing it, by Inaudi and colleagues, sets out three purposes: supporting the construction process, recording structural behavior, and enhancing bridge security, across a projected hundred-year service life. The paper is candid about the system’s institutional role, describing it as complementing routine inspection rather than replacing it, and about its origin, which was a collapse that shook public confidence in everyday infrastructure.
Continuous structural data from a major crossing, gathered for the whole life of the replacement bridge, is a richer physical feed than most digital twin demonstrations have ever had. It was not called a twin, and by the National Academies’ definition it still would not qualify, because the missing element is the return path: a validated model whose output changes what the owner does next. Bridge inspection practice and rating are treated in how bridges are inspected and rated, and the model-side techniques now being marketed for this work are examined in how AI is used in transportation infrastructure.
The test to apply
A useful question about any highway digital twin is what decision the model is authorised to change. If the answer is that engineers look at it and then decide as they would have anyway, the system is a viewer. If the answer is that a monitored measurement triggers a different inspection interval or a different posting, the system is doing what the word promises, and it then owes the public the validation record to justify that authority.
Almost nothing in the highway sector currently sits on the second side of that line, which is not a scandal. Asset databases, open model formats and instrumented structures are the prerequisites, and the sector is genuinely building them, one state manual at a time. Adopting the vocabulary of the finished thing while the prerequisites are still being assembled has a cost, and the National Academies named it: the publicity now runs ahead of the evidence, and every agency that describes its asset inventory as a twin makes the evidence harder to find.