Digital Twin

A digital twin is a virtual model of a physical asset, process, or system that stays synchronized with its real counterpart through a continuous flow of data, and that sends decisions back to it. The physical side might be a single pump, a production line, a wind farm, or a city district; the virtual side combines its geometry, its current state from sensors, and models of how it behaves, so operators can monitor it, test "what-if" changes in simulation, and predict failures before they happen. Michael Grieves introduced the concept at the University of Michigan in 2002 as a model for product lifecycle management, and NASA's John Vickers gave it the name digital twin around 2010. The ISO 23247 series, published in 2021, defines a digital twin framework for manufacturing.

Digital Model, Digital Shadow, and Digital Twin#

The term is often applied to any 3D model or dashboard, so researchers distinguish three levels by how data moves between the physical and virtual objects. A 2018 review by Kritzinger and colleagues calls a virtual model that is updated by hand a digital model, one that receives live data automatically but changes the physical asset only through people a digital shadow, and one with automatic data flow in both directions a digital twin.

A simulation, in comparison, explores what-if scenarios with no live link to a specific asset, and a digital twin runs such simulations starting from the asset's current state. 3D reconstruction techniques such as photogrammetry, neural radiance fields, and Gaussian splatting supply a twin's geometry and appearance.

How a Digital Twin Works#

A working twin connects several layers. Industrial IoT sensors, programmable logic controllers, and cameras measure the physical state, and protocols such as OPC UA and MQTT stream those readings into a data platform. Behavior models then range from physics-based simulations, such as finite-element and fluid dynamics models, to machine learning surrogates trained on simulation runs or historical data that answer in milliseconds. Analytics on top of the synchronized state cover anomaly detection, predictive modeling of remaining useful life, and time series analysis of throughput, and their results flow back as adjusted setpoints, maintenance work orders, or new production schedules.

Computer vision turns ordinary cameras into twin sensors. Object detection counts parts, pallets, and vehicles, multi-object tracking follows them through a site, pose estimation records how workers move for ergonomic studies, and segmentation measures fill levels or material piles. Ultralytics YOLO26 track mode gives each detected object a persistent ID across video frames, so a twin can update one record per forklift or pallet with its latest position instead of creating new objects on every frame. Data also flows the other way: a twin's rendered scenes produce labeled synthetic data for training those vision models, and domain randomization of textures and lighting helps them transfer to real cameras.

Real-World Applications#

• Manufacturing: Factory twins let engineers rearrange production lines, reprogram robot paths, and balance workloads virtually before touching the real line, and machine twins schedule maintenance from vibration, temperature, and current data.
• Logistics and warehousing: Warehouse twins combine camera-based tracking of goods and vehicles with order data to test layouts, slotting strategies, and robot fleets.
• Energy and aerospace: Jet engine and wind turbine twins combine physics models with sensor streams from each unit to estimate wear on individual components and time inspections accordingly.
• Smart cities: City twins model buildings, traffic, and utilities, and vehicle counts from traffic cameras calibrate their traffic simulations. Virtual Singapore, a S$73 million national program launched in 2014, models the entire city-state in 3D for urban planning studies.