Digital Twins: A Survey

Advances in computation, geographic data and AI are expanding what we can simulate—and who can use the results. Transport planners can test road closures. Utilities can examine how electric vehicles and heat pumps will change demand. Fire agencies can estimate where a wildfire will spread. Communities can explore competing proposals for the places they live.

The digital-twin industry is growing around these applications: models of real places and systems, supplied with data from the world they represent. This survey asks who uses them, which engines and platforms underpin them, what they can do today, and how far they have progressed toward whole systems modeling. A further frontier is coming into view: using models to discover ways of living and producing within ecological systems whose interactions exceed our unaided ability to reason about them.

What is a digital twin?

A digital twin is a computational representation of a real object, place or system, linked to information about its real-world counterpart. The connection allows the model to be updated as conditions change and used to explore what might happen next—or what would happen if something were changed. The update interval depends on the application: traffic signals and long-term housing plans operate on different clocks. [1]

A traffic twin, for example, combines a road network with traffic counts, signal timings and vehicle movements. It can help an operator respond to congestion or let a planner experiment with a different junction layout. Tees Valley's traffic twin, developed with Aimsun, is already used at its traffic management centre to identify delayed buses and support changes to signal operation. [2] [3]

The wider simulation field includes the design of individual machines and operator training—for example, learning to control a tree harvester safely. The focus here is large-scale systems with public consequences: cities and regions, electricity and water networks, transport and crowds, watersheds and ecosystems. Their users include corporations, governments and citizens, often with different interests in the same place.

There are several names for this part of the field. Urban and territorial digital twins describe models organized around places. Infrastructure twins describe networks and assets. Cyber-physical-social systems explicitly include people and their behavior. Multiscale describes models that connect levels of detail, such as individual households and a regional electricity grid. These terms overlap; the survey uses the plain description “large-scale digital twins” rather than treating any one of them as a universal category. [4] [5]

Who uses them, and for what?

Cities & placesMovement & logisticsEnergyWaterFood & landLiving systemsWeather & climatePeople & societyFireSensory landscape

Use	A concrete question	Systems and examples
Traffic and public transport	Where will vehicles go after a closure, and can signal changes help a late bus?	Aimsun; Tees Valley's operational traffic twin; SUMO and MATSim modeling engines. [2] [6] [7] [3]
Crowds and evacuation	How will people move through a station or leave an area during an emergency?	AnyLogic pedestrian models; GAMA evacuation models. [8] [9]
Electricity and energy transitions	Where will electrification require more network capacity, storage or generation?	Neara; PyPSA; the UK NDTP's Isle of Wight energy demonstrator. [10] [11] [12]
Flooding, rivers and water services	Which areas will flood, and how will changes to rivers, drainage or pumps affect them?	Deltares Delft3D; DHI; Xylem Vue. [13] [14] [15]
Wildfire and exposure	Where might a fire spread, which assets are exposed, and where would mitigation help?	Technosylva's operational and planning tools for utilities and fire agencies. [16]
Housing and land use	How might zoning or transport investment change development and household location?	UrbanSim; Urbanly CityCompass. [17] [18] [19]
Ecosystems and land management	How might land-use changes affect water, habitat, carbon or agricultural production?	InVEST, DSSAT and GCAM supply models for these questions; Earth Index helps map environmental features. [20] [21] [22] [23]
Weather and climate	What conditions are likely next week, and how might a region's climate change over decades?	WeatherNext, NVIDIA Earth-2 and Destination Earth. [24] [25] [26]
Sensory landscape	How do sound, light, smell and barriers affect people and other animals? Could a different street, crossing or lighting scheme make a place easier to inhabit?	Walk Score for walkability; CadnaA for noise; Radiance for light; GRAL for odor dispersion; Circuitscape for habitat connectivity. Components of a possible coupled model. Appendix G.
Public discussion and learning	What changes when we combine different climate policies or land-use proposals?	En-ROADS; Geodesignhub; public city-model and planning interfaces. [27] [28] [29]

This includes both operating twins and the simulation engines or data services used to build them. A general crop model, a traffic engine and a maintained twin of a particular region are different products, but all belong in an account of how these systems are assembled.

The audiences are visible in the applications. A utility wants to keep service running and choose investments. A local authority wants to compare transport, housing and environmental proposals. Residents may want to inspect a proposed road, explore an alternative, or understand which neighborhoods benefit. The same underlying model can support all three, provided each has an interface suited to its task.

The engines and platforms

The software market divides into geographic platforms, simulation engines and applications built for particular domains.

Esri and Bentley are central reference points for the geographic and infrastructure layer. ArcGIS organizes maps, spatial analysis and related data; Bentley iTwin integrates engineering and infrastructure information. Dassault Systèmes supplied the platform used for Virtual Singapore. These systems can give different teams a common representation of a place, its buildings and its assets. [30] [31] [32]

Simulation engines supply the behavior. AnyLogic supports agent-based, discrete-event and system-dynamics models. GAMA combines spatial data and agent-based modeling. MATSim models the travel choices and movements of many agents across transport networks; SUMO simulates vehicles, public transport and pedestrians. PyPSA provides power and energy-system modeling. Deltares and DHI supply water-modeling tools. A city project may use several of these engines. [14] [8] [9] [6] [7] [12] [15]

Specialist platforms package models around a job. Aimsun serves transport planners and traffic operators. Neara serves electricity networks. Urbanly and UrbanSim address development and land use. Technosylva combines fire-spread modeling, weather and asset information for wildfire operations and mitigation. These are often the products a buyer encounters before encountering the underlying mathematics. [10] [2] [17] [19] [16]

“Which engine?” is therefore partly a question about the model: roads, electrical networks and household decisions behave differently. The larger platform choice also concerns the data, interfaces and integrations a project needs. Appendix A separates those roles and lists representative systems.

What computation and AI are changing

Weather forecasting provides a concrete example. WeatherNext 3 uses satellite observations and training on weather-station measurements to produce hourly forecast updates, reaching five-kilometer resolution for selected variables. Results are distributed through data services and Google's products. That combination makes a computationally demanding model available to people who will never run the model themselves. [24] [33]

Earth Genome reduces a different barrier. Its Earth Index application lets a user label examples of a feature in satellite imagery, search for similar locations, inspect the results and export a map. Applications include locating mining activity and livestock facilities. In a September 2026 study, the team combined SSL4EO and AlphaEarth representations to improve cattle-facility detection. Reusable geospatial models are becoming ingredients in practical environmental investigations. [20] [34] [35]

Urbanly adds a conversational interface to urban simulation: the assistant helps set up and run experiments, while the simulation engine produces the numerical results. This changes how a planner can work with a model. It does not require the planner to build each experiment through code or a specialist modeling interface. [36]

These developments affect different costs: producing forecasts, extracting information from imagery, and setting up experiments. Together they make richer models available through ordinary data services and browser tools. Local data preparation and model calibration still account for work that a new interface cannot automatically eliminate.

Are there whole systems models yet?

There are already models that connect several domains. They differ in what they connect and how much detail they retain.

GCAM connects energy, water, agriculture and land, the economy, and climate. It can examine, for example, how a change in energy policy alters land and water demand. Developed over decades, it represents a substantial whole systems tradition that predates current digital-twin branding. Its purpose is to explore regional and global scenarios, rather than reproduce every building or household. [21]

PyPSA-Eur connects electricity to transport, buildings and industry. It can examine how an energy transition changes generation, networks, storage and demand across sectors. It is an open energy-system model, not a live operational twin of every connected asset. [37]

Urban models connect people and development. UrbanSim represents location choice and urban growth, with connections to transport modeling. Urbanly publishes comparisons between simulated and recorded housing development in parts of Los Angeles and New York. These models can address interactions that a geometric city model alone cannot: where households locate, what gets built, and how development responds to the modeled rules. [17] [18]

Connected infrastructure models address cascading failures. CReDO examines climate hazards and dependencies between networks. A flooded electricity asset can affect water pumping or communications, so the consequence extends beyond the damaged asset. Its September 2026 MVP showcase presents this approach to resilience planning; a final platform release is planned for 2028. [38] [39]

Whole systems modeling therefore exists in several forms: integrated models, linked domain models, and applications combining common data with specialist simulations. A comprehensive, routinely updated model of a city's physical, ecological, economic and social behavior remains a much larger undertaking. The examples above solve portions of that problem at different scales.

National and territorial twins

Virtual Singapore made the national-scale ambition visible: a shared 3D representation containing information about buildings, terrain and the city around them. GovTech's description illustrated uses such as assessing rooftop solar potential, flooding, wind and urban heat. Singapore's current geospatial programs continue to support applications including flood-risk assessment and solar-energy analysis. [40] [41]

The UK National Digital Twin Programme takes a broader infrastructure approach. Its published Isle of Wight demonstrators include electricity, housing retrofit and emergency planning. The energy demonstrator links network questions to electrification and renewable generation. The housing example brings together building characteristics, eligibility for retrofit funding and household vulnerability. Those examples show why social and policy information belongs alongside physical infrastructure. [11] [42] [43] [44]

These programs also expose the institutional scale of the task. A city's roads, electricity, water, land records and housing data are maintained by different organizations. Connecting their models requires agreements about sharing and updating that information, as well as software. DAFNI supplies research infrastructure for collaborative modeling; work such as Wenwen Li's at ASU addresses the discovery and integration of distributed geospatial data. [45] [46]

Making the models usable

Access already takes several forms. Helsinki publishes city-model data and a browser-based energy atlas. En-ROADS makes climate-policy scenarios available for public learning. Geodesignhub gives stakeholders a shared setting for developing and comparing spatial proposals. Earth Index puts environmental mapping tools in the hands of investigators. [20] [27] [28] [29]

A public planning interface could let someone move a proposed cycle route, compare traffic changes, and share the alternative with a council. Delivering that experience requires more than exposing an expert application's controls: the scenario must load quickly, the choices must be understandable, and the results must identify what changed. Whether residents can simply view a model or actually propose alternatives is an important product distinction.

The full stack runs from satellite imagery, sensors and administrative records through geographic data, simulation engines and linked models to the interface where somebody uses the result. A working application also needs a team to maintain those connections as roads, assets, populations and data sources change.

The industry has established suppliers for geography, transport, water and energy, alongside newer companies building environmental search, urban simulation and AI interfaces. Reusable connections between these tools would let more cities and watersheds draw on this accumulated modeling work.

Modeling beyond human intuition

The applications above often start with systems we have made explicit: a road network, an electricity grid, a zoning code, and a known set of controls. A more consequential frontier concerns systems whose organization is largely implicit. Climate, soils, water, organisms and human activity influence one another through relationships that we have only partly observed or formalized. Timothy Morton's term hyperobjects captures a related difficulty: phenomena distributed across so much space and time that we cannot encounter them as a comprehensible whole. [47]

Working with these systems means discovering more of their structure as we go. A useful model could help reveal a relationship we had overlooked, suggest an intervention we would not have considered, and identify an observation that would tell us whether it works. This extends the purpose of a twin from operating a system whose controls we understand to exploring possibilities within a system we are still learning to describe.

Agriculture makes the opportunity tangible. Uniform crops, standardized inputs and repeatable operations make production easier to organize. Machinery, labor, markets and policy all help explain that pattern. There is also a cost to managing ecological complexity: the best crop mixture for one soil, season and water regime may perform differently in another. Diversified farming already has a substantial history and evidence base. A 2020 synthesis found that diversification improved several ecosystem services and biodiversity without an average crop-yield penalty, although outcomes varied. Machine learning could extend farmers' and ecologists' ability to work with that diversity. [48]

Imagine planning a productive watershed as a changing mosaic of crops, grazing, trees, wetlands and habitat. A model could search combinations of planting, rotation, irrigation and restoration that would be impractical to compare by hand. The question becomes: which arrangements sustain food production and livelihoods while improving soil, water and biodiversity? This is a search through a fitness landscape with several objectives, delayed effects and changing conditions. There may be many useful sweet spots, and a gain in one outcome can come at another's expense.

In his essay Natural General Intelligence, Ryan Orbuch proposes a nature model that learns environmental state and dynamics across connected Earth systems, anticipates responses to intervention and helps decide what to measure next. It is a written proposal for a direction of research. The compelling idea is that greater intelligence could reveal opportunities to improve living systems that we have not yet recognized—or even known how to ask about. [49]

Fei-Fei Li's World Labs approaches world modeling through spatial intelligence. Its September 2026 Atlas announcement demonstrates scene reconstruction, generation and workflows for robotics simulation. AMD's agreement to acquire World Labs for approximately US$8.2 billion, reported by The Standard and confirmed in AMD's filing, shows the scale of investment entering this adjacent field. These developments suggest more of a model's structure can be learned from observations. Extending that capability to ecological responses over seasons and decades remains an open research challenge. [50] [51] [52]

Human intuition and machine reasoning would have complementary roles. Farmers, scientists and communities contribute experience, goals and constraints. Models propose combinations, estimate their consequences and identify where uncertainty changes the decision. Field trials and continued observation reveal which predictions survive contact with a particular place. The result could be a growing capacity to act intelligently within systems that remain too complex for any one person to understand in full. Appendix F follows this loop through the stack.

The goal is to leave a landscape with greater capacity to support life: healthier soils, more habitat, cleaner water and durable livelihoods. Carbon-negative activity belongs in that ambition, with emissions and removals considered across the whole activity rather than only within a field boundary. Living systems already regenerate many of the conditions on which production depends. Whole systems models could help us strengthen that regenerative capacity—and make its improvement an explicit measure of success.