The science behind digital twins

01 FROM APOLLO TO THE DIGITAL THREAD

The concept did not begin with a name. During the 1960s, NASA used paired physical and simulated models of the Apollo spacecraft to diagnose failures in real time. When Apollo 13's oxygen tank ruptured, engineers on the ground used a growing digital replica of the vehicle to test recovery procedures before radioing them to the crew. The replica was not called a digital twin, but it was one in every functional sense: a computational model that mirrored a physical system and stayed updated with live data from that system.

The formal concept arrived decades later. David Gelernter's 1991 book Mirror Worlds envisioned computational copies of physical systems that citizens could interact with. In 2010, NASA engineer John Vickers and Michael Grieves gave the idea its first practical definition in the context of spacecraft life-cycle management. They identified three parts: the physical system, the digital representation, and the communication channel linking them. That channel is now called the digital thread, and it is what separates a digital twin from an ordinary simulation.

A simulation runs on assumptions. A digital twin runs on measurements. The difference is the thread. If the data link breaks, the twin becomes a model, not a twin. If the model cannot ingest and react to new data, it cannot predict the physical system's future state. The thread is not a cable. It is a protocol, a data pipeline, and a synchronization strategy that keeps the virtual and physical worlds aligned within a useful time window.

02 WHAT MAKES A MODEL A TWIN

A digital twin is a computational model that emulates the behavior of a physical system and receives real-time data from that system to update itself along its life cycle. The strict definition requires continuous synchronization. The twin does not merely represent the physical system at one moment in time. It tracks the physical system as it changes, adjusting its internal state to match.

Three subtypes are commonly distinguished. A digital twin prototype captures the designs, analyses, and processes that define a product before it exists. A digital twin instance is the twin of one specific manufactured unit, linked to its physical counterpart for the remainder of that counterpart's life. A digital twin aggregate is the collection of all instances of a product, whose pooled data enables population-level analysis, prognostics, and learning. The specific information contained in each twin is driven by use cases. The twin is a logical construct, meaning the actual data and information may be contained in other applications.

Not every virtual model qualifies. Researchers at the United Kingdom's National Physical Laboratory have argued that overly broad definitions, where any executable virtual model counts as a twin, risk reducing the term to a buzzword. A model that does not produce results equivalent to measured quantities, or that is not dynamically updated in accordance with those measurements, is a simulation, not a twin. The distinction matters because the twin's value comes from its live coupling, not from its geometry.

03 THE PHYSICS ENGINE INSIDE THE TWIN

Most digital twins are built on physics-based models. A jet engine twin, for instance, encodes thermodynamic equations that govern compressors, combustors, and turbines. A structural twin encodes finite element equations that describe stress, strain, and fatigue. A fluid dynamics twin encodes Navier-Stokes equations that predict flow, heat transfer, and pressure loss. These models are computationally expensive, and the twin must run them fast enough to stay ahead of the physical system's evolution.

The fidelity of the physics model determines what the twin can predict. A low-fidelity model might capture bulk temperature trends but miss localized hot spots that cause premature wear. A high-fidelity model might resolve individual blade-level heat transfer but require hours of computation for minutes of simulated time. The designer trades fidelity against speed, and that trade-off depends on the twin's purpose. A maintenance twin needs to predict failures weeks in advance; a control twin needs to predict state changes in milliseconds.

Some twins use reduced-order models, which compress the full physics equations into a smaller surrogate that runs orders of magnitude faster while preserving accuracy within a defined operating envelope. Others use machine learning to learn the input-output mapping of the full physics model, replacing the expensive computation with a neural network inference. The most advanced twins fuse both approaches, using the physics model where the operating conditions are well understood and the learned model where the physics is too complex or too slow to compute directly.

04 THE DATA PIPELINE THAT KEEPS THE TWIN ALIVE

The digital thread is not a passive wire. It is an engineered data pipeline that handles acquisition, transmission, processing, and ingestion. On the physical side, sensors capture temperature, pressure, vibration, strain, flow rate, or any quantity the twin needs to track. These measurements are timestamped and transmitted, often wirelessly, to a processing layer that cleans, calibrates, and aligns them before feeding them to the twin.

Latency is a first-order design constraint. If the twin receives data from the physical system seconds after the event occurred, its internal state lags behind reality. For slow-changing systems like a building's energy consumption, a lag of minutes may be acceptable. For fast-changing systems like a jet engine's turbine blades, the twin may need data within milliseconds to remain useful. The pipeline must match the dynamics of the system it mirrors.

Data quality is the other constraint. Sensors drift, calibrate, and fail. The pipeline must detect anomalous readings, reject corrupted packets, and interpolate missing data without introducing errors that propagate through the twin. A twin that trusts a broken sensor will diverge from its physical counterpart, and once it diverges, its predictions degrade. The pipeline is the twin's immune system.

05 STATE ESTIMATION AND THE KALMAN FILTER

A twin does not receive a perfect description of its physical counterpart's state. It receives noisy, partial measurements from a limited set of sensors. To maintain an accurate internal state, the twin must estimate the full state from those measurements, and it must do so continuously. The mathematical tool that makes this possible is the Kalman filter and its extensions.

A Kalman filter fuses two sources of information: the physics model's prediction of the next state and the sensor's measurement of the current state. Each source has uncertainty, and the filter weights them according to their relative confidence. If the sensors are precise, the filter trusts the measurements. If the model is precise, the filter trusts the prediction. The result is a state estimate that is better than either source alone.

For nonlinear systems, the extended Kalman filter linearizes the model around the current estimate. The unscented Kalman filter propagates a set of carefully chosen sample points through the nonlinear model to avoid linearization errors. Particle filters represent the state as a cloud of hypotheses, each propagated and weighted by how well it explains the measurements. These are the same algorithms used in robotics for localization and tracking, and they are the mathematical backbone of a twin's ability to stay synchronized with a physical system it cannot fully observe.

06 UNCERTAINTY QUANTIFICATION

A twin that produces a single number as a prediction is incomplete. Every prediction carries uncertainty, and the science of digital twins demands that uncertainty be quantified, propagated, and communicated. A twin that predicts a component will fail in 30 days is useless if it cannot say whether that estimate is accurate to a day, a week, or a month.

Uncertainty enters the twin from three sources. Model form uncertainty arises because the physics equations are approximations of reality. Parameter uncertainty arises because the model's coefficients are not known exactly. Observation uncertainty arises because the sensors are noisy and incomplete. A rigorous twin quantifies all three and propagates them through the computation so that every prediction arrives with a confidence interval.

Methods for uncertainty quantification range from Monte Carlo sampling, which runs the model many times with varied inputs to build a distribution of outputs, to polynomial chaos expansion, which represents the output as a function of the uncertain inputs. For expensive physics models, surrogate-based methods use a fast learned model to approximate the full model's output distribution. The goal is always the same: a twin that tells you not just what will happen, but how confident it is in that prediction.

07 WHEN THE TWIN BECOME A PROGNOSTIC ENGINE

The most valuable twin does not just mirror the present. It predicts the future. Prognostics is the science of estimating remaining useful life, and it is where the twin's physics model, data pipeline, state estimation, and uncertainty quantification converge into a single capability.

A prognostic twin tracks the physical system's degradation state, estimates the current level of wear or damage, and projects forward to predict when that damage will reach a failure threshold. The projection uses the physics model to simulate future operating conditions and their effect on the degradation variable. The uncertainty in the current state estimate and in the future operating conditions propagates into the remaining-useful-life prediction, producing a distribution rather than a single number.

This is where digital twins deliver their hardest economic value. A jet engine twin that predicts a turbine disk will fail in 200 cycles, plus or minus 20, allows the airline to schedule maintenance during a planned downtime rather than experiencing an unscheduled removal. The twin does not prevent the failure. It converts an unexpected failure into a scheduled event, and that conversion is where the twin earns its cost.