Digital twin
The Model That Predicts the Failure Before the Hardware Does: Digital Twins and Simulation-First Validation
The prototype costs real money and weeks, and the failure it hides costs a hundred times more to find in the field than at validation. A digital twin is not a 3D model, it is a model that behaves like the physical system, fed by data from it, whose answers change the physical system. Example Labs' twin of its sensor hub predicted the rail sag that killed four of ten hubs in the field, as a band that straddled the sensor's minimum voltage, which is exactly what 'no margin' looks like before the hardware is built.
The prototype costs real money and weeks. The board costs money to spin, the parts cost money to buy, the assembly costs money to build, and the test costs money to run, and when the prototype fails, the team learns one thing: that the prototype fails. It does not learn why, and it does not learn what else would have failed, because the prototype can only fail one way at a time.
A digital twin is the thing that fails faster, cheaper, and in every way at once, before the hardware exists. It is not a 3D model. It is not a simulation you run once and file away. It is a model that behaves like the physical system, fed by data from the physical system, whose answers change the physical system. And the difference between a twin and a simulation is the difference between a product that ships and a product that fails in the field.
What a digital twin is, and what it is not
A digital twin is a computational model of an intended or actual real-world physical product, system, or process, a physical twin, that serves as a digital counterpart of it for purposes such as simulation, integration, testing, monitoring, and maintenance. By its strict definition, a digital twin is distinguished from an ordinary simulation in that it continuously uses real data from its physical counterpart to dynamically synchronize with the real system.
A model that operates without data from its physical counterpart may also be described as a digital twin, but this is considered an overly broad and largely marketing-oriented interpretation of the concept.
Three things have to be true, and the third is the one that separates a twin from a picture:
- It is a model that behaves like the physical entity it represents. Not a model that looks like it. A model that behaves like it, the same inputs produce the same outputs, within a known band.
- Data from the physical entity is used to maintain the accuracy of the model. The twin is fed by the real system, sensor readings, test results, field logs, and it gets better as the real system runs.
- Information from the model is used to impact the physical entity. The twin's answers change the real system: a design change, a test that gets run, a limit that gets set.
Any application that includes all three attributes can be considered a digital twin. A 3D CAD model has the first and neither of the other two, it is a blueprint, not a twin. A simulation run once has the first and neither of the other two, it is a calculation, not a twin.
And a dashboard that shows sensor data has the second and neither of the other two, it is a monitor, not a twin. The twin is the loop: the model predicts, the physical system measures, the measurement corrects the model, and the corrected model predicts better.
The history is older than the name. The concept can be traced back to the early days of the space age: in the 1960s, NASA pioneered the idea of a digital twin to simulate spacecraft and troubleshoot issues in real time. After Apollo 13's oxygen tank exploded in April 1970, the agency's engineers combined spacecraft simulators with real-time data from the physical ship to evaluate the situation and form an appropriate response.
The term itself was formulated much later, in 2002, by Michael Grieves in the context of product lifecycle management to guide industrial design and management. By around 2010, NASA had adopted digital twins as detailed virtual models of physical assets, emphasising that a true digital twin comprises three interconnected elements: a physical entity, a virtual model, and the data connections that enable them to mirror each other in real time.
It is the dynamic, bidirectional link, sensor data flowing into the model and simulation insights flowing back to inform the physical system, that elevates a digital twin beyond a static blueprint or 3D model.
The manufacturing standard is ISO 23247, which defines a digital twin as a "fit for purpose digital representation of an observable manufacturing element with a means to enable convergence between the element and its digital representation at an appropriate rate of synchronization." Note the words "fit for purpose" and "at an appropriate rate", the twin does not have to be perfect, and it does not have to be real-time. It has to be good enough for the decision it supports, and synchronised often enough for that decision.
A twin that predicts a rail sag to within 0.05 V is fit for the purpose of deciding whether the rail has margin. A twin that predicts it to within 0.5 V is not.
There are three subtypes, and the distinction matters for a team deciding what to build. A digital twin prototype (DTP) is the twin of a product that does not exist yet, the twin that predicts the failure before the hardware does, which is this article. A digital twin instance (DTI) is the twin of one specific physical unit, fed by that unit's data, the twin that follows one hub through its life.
And a digital twin aggregate (DTA) is the collection of instances, the twin that learns from the whole fleet. A team that builds the prototype twin first gets the DTI and the DTA for free later, because the prototype twin becomes the instance twin the day the first unit ships.
The ladder
The twin does not replace testing. It organises it. The automotive industry has a ladder for that proof, Model-in-the-Loop, Software-in-the-Loop, Processor-in-the-Loop, and Hardware-in-the-Loop, usually shortened to MIL, SIL, PIL, and HIL. Each rung adds a bit more reality, and each one exists because the rung below it is blind to a certain class of bug. The rule behind the ladder is simple: find a bug at the lowest rung that can see it.
Every modern car is a network of 70 to 150 electronic control units, and every one of them runs software that has to be proven safe before it ships. The ladder is how the industry proves it, and the same ladder works for a sensor hub, a cobot cell, or any system with a controller and a plant. The V-model maps development activities on the left side of the V, requirements, design, coding, to corresponding verification and validation activities on the right side, with simulation-based testing bridging the two.
Each stage inherits the validated artefacts of the previous one, progressively reducing the gap between simulation and physical reality.
MIL, Model-in-the-Loop. Both the control algorithm and the plant model run as high-level models, without compiled code or physical hardware. The controller is a model, a Simulink model, or the Python simulation from Article #3, and the plant is a model too: the motor, the battery, the arm, the sensor. The loop is closed: the controller's output goes into the plant model, and the plant's reaction comes back as the next input. MIL is the earliest practical test stage, and the fastest, you are not compiling or deploying code. What it finds: wrong control logic, wrong requirements interpretation, unstable control loops. What it misses: anything about code, timing, or hardware. The PID loop from Article #3 was tuned at this rung: P alone leaves a steady-state error of a third of the step, forever, and the simulation showed it before anything moved.
SIL, Software-in-the-Loop. The model is replaced by the production code, compiled and executed on a host PC rather than on the target hardware, against the same plant model. SIL is the fastest and most cost-efficient method for verifying algorithmic correctness before the hardware is manufactured, and it runs in CI, a regression suite that takes minutes in SIL can take hours or days on HIL. What it finds: errors introduced by code generation or hand coding, integer overflow and underflow, type casting issues, fixed-point arithmetic errors, wrong state machines. What it misses: compiler and processor effects, real-time behaviour. The Kalman filter from Article #4 belongs here: the filter's arithmetic is tested on the host, against the plant model, before it ever runs on the STM32H7.
PIL, Processor-in-the-Loop. The code, compiled with the real target compiler, executes on the target processor, an evaluation board or an instruction-set simulator, while the plant still runs on the PC. What it finds: effects of the target compiler and processor, fixed-point arithmetic, stack use, execution time. What it misses: I/O circuits, electrical faults, real bus timing. A common technique between levels is back-to-back testing: run the same test inputs through MIL, SIL, and PIL and compare the outputs. If the outputs differ, the difference is a bug in the code generation, the compiler, or the processor, and the rung that found it tells the team where to look.
HIL, Hardware-in-the-Loop. The finished ECU, hardware, software, connectors, wired to a real-time simulator that pretends to be the rest of the system: the sensors, the actuators, the vehicle, the battery. HIL is the final pre-vehicle simulation stage, and the first stage where the full communication stack can be tested. What it finds: I/O and electrical behaviour, diagnostics and fault codes, real CAN, LIN, and Ethernet timing, fault reactions, start-up and shutdown. What it misses: real vehicle dynamics and environment, which only vehicle tests show. And the catch: HIL benches are expensive, there are never enough of them, they exist only after the ECU hardware exists, and a regression suite that takes minutes in SIL can take hours or days on HIL. That is why the ladder exists, so the team does not wait for the HIL bench to find the bugs the lower rungs could have caught.
The ladder is also the twin's construction plan. The MIL model becomes the twin's core. The SIL stage proves the code that runs inside the twin. The PIL stage proves it on the target processor. The HIL stage proves the hardware against the twin. And when the hardware ships, the twin keeps running, fed by the hardware's data, predicting the hardware's failures. The twin is the ladder, kept alive after the product ships.
VVUQ: how to know whether the twin is telling the truth
"All models are wrong, some are useful." That is George E. P.
Box, and it is the first line of every honest discussion of simulation. All models rely on assumptions and approximations in the formulation of the model. All models require input data to describe the system and its surroundings, which are approximate, possibly not measurable, or not knowable in advance.
All models require numerical solution, which results in numerical errors. The question is not whether the model is wrong. The question is whether it is useful, and whether the team knows how wrong it is.
ASME's answer is VVUQ: verification, validation, and uncertainty quantification. The ASME committee on verification and validation was formed in 2001; the first edition of ASME V&V 10, Guide for Verification and Validation in Computational Solid Mechanics, was published in 2006; and the committee's name was changed in 2021 from V&V to VVUQ, to highlight the importance of uncertainty quantification. The three assessments are:
Verification, did we solve the equations correctly? Verification is the assessment of the accuracy of the numerical solution relative to the exact solution to the mathematical model. In verification, the relationship of the simulation to the real world is not an issue. The code is checked against the equations: the Method of Manufactured Solutions is the gold standard for code verification, and mesh convergence studies quantify the numerical error. Verification answers the question, "Did we build the model right?"
Validation, did we build the right model? Validation is the assessment of the accuracy of the simulation results relative to experimental measurements of system responses. In validation, the relationship between computation and the real world, experimental data, is the issue. The twin's prediction is compared against a measurement: the predicted rail sag against the measured rail sag, the predicted settling time against the measured settling time. Validation answers the question, "Did we build the right model?"
Uncertainty quantification, how wrong could the answer be? Uncertainty quantification is the estimation of the total uncertainty in the simulation result due to all sources of error and uncertainty in the modeling and simulation process. The twin does not report a number. It reports a band. The cell's internal resistance varies by lot and by temperature, and it is not knowable in advance, so the twin's rail sag is a band, and the band is the answer. If the band straddles the limit, the design has no margin, and the twin has done its job.
And the order matters: verification must precede validation. You cannot validate an unverified code; doing so merely validates the buggy implementation. The twin that predicts the right answer for the wrong reason is worse than no twin at all, because the team trusts it.
The infamous failures are the warning. ASME's own VVUQ teaching material lists them: the Space Shuttle Columbia disaster, the Fukushima nuclear disaster, both cases where erroneous modeling contributed to a system failure. The twin is not a guarantee. It is a discipline: verify the code, validate the model, quantify the uncertainty, and never trust a number without its band.
The twin and the simulation
Simulations are closely related to digital twins but differ in two ways. Both involve digitally replicating a system and its processes, and simulations can be run on digital twins to investigate potential scenarios and predict outcomes. However, simulations are not usually connected to an outside entity, and they lack real-time data updates, two core components of digital twins.
That distinction decides what the team builds. A simulation answers a question once: will the S-curve settle faster than the trapezoidal profile? Article #29 answered it in simulation, the S-curve commanded 1.058 s against the trapezoidal 0.857 s, but settled at 1.058 s against 3.042 s, 65.2% faster overall, with a residual of 0.038° against 0.734°, 19.2× smaller, and the answer was final.
A twin answers a question continuously: is the arm still settling in 1.058 s, or has the joint worn, the load changed, the temperature drifted? The twin watches the real arm, compares its motion to the simulated motion, and flags the day the two diverge. The simulation proves the design.
The twin watches the product.
And the twin answers two questions that sensor monitoring alone cannot: what is causing this outcome, and what would happen if we changed this parameter? A sensor tells the team the rail sagged. The twin tells the team the rail sagged because the cell aged, and that replacing the cell with a larger one, or raising the brownout threshold, would have kept the rail above the sensor's minimum. That is the twin's value: not the data, but the counterfactual.
What it costs, and what it returns
The returns are real, and they are also the part of this article that needs the most care, because the numbers come from vendors and industry surveys, not from measurements the team can check. McKinsey's research documents digital twin programs cutting development time by up to 50%, reducing transportation and labor costs by up to 10%, and increasing revenue by as much as 10% for organizations building customer-facing digital twins.
BMW's virtual factory mirrors more than 30 of BMW's global production sites in real time, and according to BMW Group itself, it has reduced production planning costs by up to 30%. In advanced industries, survey data indicate that almost 75 percent of companies have already adopted digital-twin technologies that have achieved at least medium levels of complexity. NIST's Economics of Digital Twins research puts the total potential impact of digital twin adoption in U.S. manufacturing at $37.9 billion, against a documented $193.6 billion in combined maintenance costs and operational losses across the discrete manufacturing sector it examined.
The market-size figures, by contrast, are not worth citing, and the reason is instructive. One research firm values the digital twin manufacturing market at $3.6 billion in 2024, projected to $42.6 billion by 2034. Another values the global market at about $34 billion in 2026. A third values the digital twin in manufacturing market at $8.12 billion in 2025, projected to $139.57 billion by 2035. Three firms, three definitions, three answers that differ by a factor of four in the same year.
The market is growing fast, every firm agrees on that, but the number depends on what the firm counts as a digital twin, and most of them count simulations, dashboards, and 3D models too. The article does not cite a single market size as fact, because the spread is the fact.
The honest way to price a twin is the way Article #40 priced a post-mortem: against the failure it prevents. A failure found at engineering validation costs 1×; at the prototype stage, 5×; at pilot production, 20×; at mass production, 50×; at field deployment, 100× or more. The twin is built at the validation stage. It pays for itself the first time it catches a failure that would otherwise have been found in the field.
The decision test
Before the twin is built, five questions. Each one has an answer that can be checked, and the twin is not started until all five are answered.
1. What is the decision the twin supports? A twin without a decision is a hobby. "Will the rail hold up under load at end of life?" is a decision. "Will the arm settle in time with the new payload?" is a decision. "Let's build a digital twin" is not a decision, and the twin built for it will answer nothing.
2. What is the model, and what is its band? The twin is a model with a known uncertainty. The cell's internal resistance varies by lot and temperature, so the rail sag is a band, not a number. If the team cannot state the band, it does not have a twin. It has a guess with a picture.
3. Which rung catches which bug? MIL catches the logic. SIL catches the code. PIL catches the compiler and the processor. HIL catches the timing and the I/O. The team that runs everything on HIL wastes the bench; the team that runs nothing below HIL misses the bugs the lower rungs would have caught for free.
4. What is the data link? The twin needs data from the physical system, sensor readings, test results, field logs, and a rate of synchronisation appropriate to the decision. A twin that predicts a rail sag does not need real-time data; it needs the field logs. A twin that watches a running arm does. "Fit for purpose" and "at an appropriate rate" are the standard's words, and they are the right ones.
5. What is the verification, and what is the validation? Verification first: does the code solve the equations correctly? Then validation: does the model match the measurements? A twin that has never been validated against a measurement is a simulation with a live feed, and a simulation with a live feed is still a simulation.
Objections
"A twin is a big-company tool. We are sixty people." The ladder scales down. MIL is a Python simulation, Article #3 ran one. SIL is the compiled firmware on a host PC, the team already has the PC. PIL is the code on the evaluation board, the team already has the board. HIL is the only rung that costs real money, and the team does not need it until the hardware exists. The twin for a sixty-person company is the MIL and SIL stages, kept alive after the product ships, fed by the field logs. That is a twin. It is not a smaller version of BMW's. It is the same thing, at the scale the decision needs.
"Our product is too simple for a twin." The sensor hub is a board, a battery, a radio, and a sensor, and the twin would have caught the failure that killed four of ten hubs in the field. Simple products fail simply, and simply is still a failure. The twin's size follows the decision, not the product.
"The model will never be accurate enough." It does not have to be accurate. It has to be accurate enough for the decision, "fit for purpose", and it has to report its band. A twin that predicts the rail sag as 0.30–0.40 V, when the sensor's minimum is 2.70 V and the rail is 3.0 V, has told the team everything it needs to know: the band straddles the limit, and the design has no margin. The twin does not need to be right. It needs to be honest about how wrong it could be.
"We will build the twin after the product ships." The twin's value is highest before the hardware exists, the digital twin prototype, the twin that predicts the failure before the hardware does. After the product ships, the twin is still valuable, it watches the fleet, and it learns, but the failures it would have prevented have already happened. Build the prototype twin first. It becomes the instance twin the day the first unit ships.
"Simulation is enough." A simulation answers a question once. A twin answers it continuously, and it answers the counterfactual: what would happen if we changed this parameter? The simulation proves the design. The twin watches the product. The team that stops at simulation has proven the design once, and then the product changes, the cell ages, the temperature drops, and the proof is out of date.
Example Labs is a made-up company we use for worked examples.
DEMO
The firm is made up. Call it Example Labs again: 60 staff, and now a twin, of the field sensor hub from Articles #5, #6, #7, #17, #27, #37, and #40 (a 100 × 60 × 40 mm enclosure, a 4-layer board, 100 × 60 mm, the event camera from Article #2, the IMU from Article #4, the STM32H7 running the Kalman filter and the PID loop, the BLE radio at 2.4 GHz, a CR2032, 230 mAh = 2,484 J), and of the UR5e-class cobot cell from Article #39 (5 kg payload, 850 mm reach, 1 m/s tool speed).
The demo's arithmetic is the article's own, verified before writing, on the published figures cited above; the firm, the hub, the twin, and the measurements are the worked example, not a measured system.
Step 1, the ladder, on the hub's firmware. The hub's firmware goes through the four rungs, and each rung catches a class of bug. At MIL, the PID loop from Article #3 runs in simulation against the plant model, and the simulation shows what Article #3 measured: P alone leaves a steady-state error of a third of the step, forever. The logic is fixed before any code exists. At SIL, the compiled firmware runs on the host PC against the same plant model, and the Kalman filter from Article #4 overflows a 16-bit accumulator on the third hour of simulated run time, because the filter's covariance grows without bound. The code is fixed before any hardware exists. At PIL, the code runs on the STM32H7, and the fixed-point arithmetic in the filter loses a bit of precision at the coldest temperature, because the target compiler rounds differently from the host compiler. The back-to-back test, the same inputs through MIL, SIL, and PIL, the outputs compared, finds it. At HIL, the real hub board is wired to a real-time simulator pretending to be the battery and the sensors, and the rail sags under the BLE transmit load, the sensor latches SDA low, and the hub stops transmitting. The bug that killed four of ten hubs in the field, in Article #40, is caught at HIL, before the hardware ships, before the field, before the 100× cost. Each rung catches the bug the rung below it cannot see, and the rule holds: find the bug at the lowest rung that can see it.
Step 2, the twin, not the simulation. The ladder is a test plan. The twin is the ladder kept alive. The twin of the hub is the MIL plant model, the battery, the rail, the sensor, the radio, plus the live data link: the field hubs' rail-voltage logs, the temperature logs, the state-of-charge estimates. The simulation predicts the rail sag once, for a fresh cell at room temperature. The twin predicts it continuously, for the cell as it ages, at the temperature the hub is actually in. And the twin answers the counterfactual the simulation cannot: what would happen if the cell were replaced with a larger one, or if the brownout threshold were raised above the sensor's latch-up voltage? The simulation proves the design. The twin watches the product, and the day the twin's prediction and the hub's measurement diverge, the twin has found the next failure.
Step 3, VVUQ, on the rail sag. Verification first: the twin's battery model is checked against the equations, does the code solve the model correctly? The Method of Manufactured Solutions is the gold standard, and the mesh convergence study quantifies the numerical error. Then validation: the twin's prediction is compared against a measurement. The twin predicts the rail sag under a BLE transmit burst, at 50% state of charge, in the cold, as a band: 0.30–0.40 V, the band is the uncertainty, because the cell's internal resistance varies by lot and by temperature, and it is not knowable in advance. The rail is nominally 3.0 V. The sensor's minimum operating voltage is 2.7 V. At the low end of the band, the rail is 2.70 V, exactly at the limit. At the high end, 2.60 V, below it. The measurement, on an aged cell in the cold, is 0.36 V, the rail falls to 2.64 V, 0.06 V below the sensor's minimum, inside the band, and the sensor latches. The twin did not say "it will fail." It said "the band straddles the limit, and the design has no margin", which is exactly what "no margin" looks like before the hardware is built. That is the twin's answer, and it is the answer the team needed: size the rail for the end-of-life cell, raise the brownout threshold, and test on an aged cell. The same fix Article #40 reached by post-mortem, reached here before the hardware existed.
Step 4, the cost, and the payback. The twin costs 5 engineer-days to build at the MIL and SIL stages, the article's own construction, labelled. The field failure it prevents cost 100 engineer-days or more, Article #40's escalation, from the Semishops multipliers: a fix that costs 1 engineer-day at validation costs 100 or more in the field. The twin pays for itself 20 times over the first time it catches a failure that would otherwise have been found in the field. And the industry figures, labelled as vendor and survey figures: McKinsey's digital twin programs cutting development time by up to 50%, a 12-week development becomes 6 weeks; BMW's production planning costs reduced by up to 30%, a $100,000 planning budget becomes $70,000. The twin is not free. It is cheaper than the failure.
Step 5, the decision. Build the twin at the MIL and SIL stages first, the cheapest rungs, the ones that catch the logic and the code. Add PIL when the code runs on the target processor, and HIL when the hardware exists. Connect the live data link only after the first field deployment, when there is data to feed it, the prototype twin becomes the instance twin the day the first unit ships, and the aggregate twin learns from the fleet. And the twin's first decision is the one in Step 3: the rail has no margin at end of life, so the rail is sized for the end-of-life cell before the board is spun. The twin is not built for its own sake. It is built for that decision.
Sources
- Nature, Scientific Reports, "Digital twins: past, present and future" (nature.com, checked 2026-11-20)
What we used
"The concept of the 'digital twin', which is a virtual representation of a physical system, can be traced back to the early days of the space age. In the 1960s, NASA pioneered the idea of a digital twin to simulate spacecraft and troubleshoot issues in real time. The term 'digital twin' itself was formulated much later, in 2002, by Michael Grieves in the context of product lifecycle management to guide industrial design and management. By around 2010, NASA had adopted digital twins as detailed virtual models of physical assets, emphasising that a true digital twin comprises three interconnected elements: a physical entity, a virtual model and the data connections that enable them to mirror each other in real time."; "It is the dynamic, bidirectional link, where sensor data flows into the model and simulation insights flow back to inform the physical system, that elevates a digital twin beyond a static blueprint or 3D model."
- Wikipedia, "Digital twin" (en.wikipedia.org, checked 2026-11-20)
What we used
"A digital twin is a computational model of an intended or actual real-world physical product, system, or process (a physical twin) that serves as a digital counterpart of it for purposes such as simulation, integration, testing, monitoring, and maintenance. By its strict definition, a digital twin is distinguished from an ordinary simulation in that it continuously uses real data from its physical counterpart to dynamically synchronize with the real system."; "A model that operates without data from its physical counterpart may also be described as a digital twin, but this is considered an overly broad and largely marketing-oriented interpretation of the concept."; "The first practical definition of a digital twin originated from NASA in an attempt to improve the physical-model simulation of spacecraft in 2010."; "Digital twins are commonly divided into the three subtypes: digital twin prototype (DTP), digital twin instance (DTI) [and digital twin aggregate (DTA)]."; "The broader idea that became the digital twin concept was anticipated by David Gelernter's 1991 book Mirror Worlds. Dr. Michael Grieves first introduced the formal three-part concept of the digital twin back in 2002 during an industry presentation centered around Product life-cycle management."; "The digital twin concept consists of three distinct parts: the physical object or process and its physical environment, the digital representation of the object or process, and the communication channel between the physical and virtual representations."
- Britannica, "Digital twin" (britannica.com, checked 2026-11-20)
What we used
"digital twin, virtual representation of a physical object, structure, or system that incorporates real-time data from the entity that it represents. Digital twins can be used to analyze performance, to conduct simulations, and to identify potential problems with their physical counterparts."; "Simulations are closely related to digital twins but differ in many ways. Both involve digitally replicating a system and its processes, and simulations can be run on digital twins to investigate potential scenarios and predict outcomes. However, simulations are not usually connected to an outside entity, and they lack real-time data updates, two core components of digital twins."; "After Apollo 13's oxygen tank exploded in April 1970, the agency's engineers combined spacecraft simulators with real-time data from the physical ship to evaluate the situation and form an appropriate response."
- Automation.com, "Defining Digital Twin: History, Attributes and Examples" (automation.com, checked 2026-11-20)
What we used
"The attributes of 'digital twin' required the following attributes: 1. It is a model that behaves like the physical entity it represents. 2. Data from the physical entity is used to maintain the accuracy of the model. 3. Information from the model is used to impact the physical entity. Any application that includes all those attributes can be considered a digital twin."; "Many regard the ISO 23247 series of standards to be the definitive source for digital twins. The definition used in 23247 is: 3.2.2 <manufacturing> fit for purpose digital representation of an observable manufacturing element with a means to enable convergence between the element and its digital representation at an appropriate rate of synchronization."; "Regarding Attribute 2, notice there is no requirement that the data is provided through a realtime connection to the physical entity. If historical data is obtained offline and used for model training through supervised learning, it can still be a digital twin."; "NASA engineer John Vickers named this concept 'digital twin' in a 2010 NASA technology roadmap."
- PatSnap, "Hardware-in-the-Loop vs Software-in-the-Loop Testing" (eureka.patsnap.com, checked 2026-11-20)
What we used
"HIL (Hardware-in-the-Loop) and SIL (Software-in-the-Loop) testing are sequential, not competing, stages in the automotive V-model development cycle, each addressing a distinct layer of the validation problem for autonomous vehicle control systems. The V-model maps development activities on the left side of the V (requirements, design, coding) to corresponding verification and validation activities on the right side, with simulation-based testing bridging the two."; "The progression through the V-model begins with MIL (Model-in-the-Loop), where both the control algorithm and the plant model run as high-level models without compiled code or physical hardware. SIL follows, introducing compiled software running on a host computer. PIL (Processor-in-the-Loop) then moves that compiled code onto the target processor in isolation. HIL represents the final pre-vehicle simulation stage, connecting the complete ECU to a real-time simulated environment. Each stage inherits the validated artefacts of the previous one, progressively reducing the gap between simulation and physical reality."
- DBC Utility, "HIL, SIL, MIL and PIL testing explained for ECU and CAN engineers" (dbcutility.com, checked 2026-11-20)
What we used
"Each level replaces less of the real system with simulation. MIL tests the model, SIL the compiled code on a PC, PIL the code on the real processor, and HIL the finished ECU with its real inputs, outputs and buses. Earlier levels are fast and cheap and catch logic errors; HIL catches what only real hardware and real bus timing reveal."; the four-level table: MIL, the controller model, the plant simulated, runs on a PC in simulation time; SIL, the controller code compiled for the PC, the plant and the ECU's surroundings simulated, a PC as fast as possible; PIL, the code compiled for and running on the target processor, the plant on a PC; HIL, the finished ECU (hardware, software, connectors), sensors, actuators, and the rest of the vehicle simulated in real time, a real-time simulator wired to the ECU; "The loop is closed at every level: the controller's output goes into the plant model, and the plant's reaction comes back as the next input."; what each level finds: MIL, wrong control logic, wrong requirements interpretation, unstable control loops (misses anything about code, timing, or hardware); SIL, errors introduced by code generation or hand coding, integer overflow, wrong state machines (misses compiler and processor effects, real-time behaviour); PIL, effects of the target compiler and processor: fixed-point arithmetic, stack use, execution time (misses I/O circuits, electrical faults, real bus timing); HIL, I/O and electrical behaviour, diagnostics and DTCs, real CAN, LIN and Ethernet timing, fault reactions, start-up and shutdown (misses real vehicle dynamics and environment); "A common technique between levels is back-to-back testing: run the same test inputs through MIL, SIL and PIL and compare the outputs."
- ErrBits, "MIL, SIL, PIL, HIL Testing, Where SystemC Virtual ECUs Fit" (errbits.com, checked 2026-11-20)
What we used
"Every modern car is a network of 70 to 150 electronic control units, and every one of them runs software that has to be proven safe before it ships."; "The rule behind the ladder is simple: find a bug at the lowest rung that can see it."; "The catch: HIL benches are expensive, there are never enough of them, they exist only after the ECU hardware exists, and a regression suite that takes minutes in SIL can take hours or days on HIL."
- AskUI, "HIL vs SIL vs MIL: The Full Testing Hierarchy" (askui.com, checked 2026-11-20)
What we used
"A bug caught in Model-in-the-Loop costs almost nothing to fix."; "The question is not whether to use MIL, SIL, or HIL. The question is which stage is the right gate for which class of defect."; "MIL, SIL, and HIL are three distinct stages in the automotive V-Model testing hierarchy, each operating at a different level of hardware fidelity."
- ASME, "Simulation Credibility Assessment: Introduction to Verification, Validation, and Uncertainty Quantification (VVUQ)" (asme.org, checked 2026-11-20)
What we used
"All models are wrong, some are useful.", Prof. George E. P. Box; "All models rely on assumptions and approximations in the formulation of the model; All models require input data to describe the system of interest and its surroundings, which are approximate, possibly not measurable, or not knowable in advance; All models require numerical solution, which results in numerical errors"; "Infamous system failures due to erroneous modeling: Space Shuttle Columbia Disaster, Fukushima Nuclear Disaster"; "Simulation verification: assessment of the accuracy of the numerical solution relative to the exact solution to the mathematical model; Model validation: assessment of the accuracy of the simulation results relative to experimental measurements of system responses; Uncertainty quantification: estimation of the total uncertainty in the simulation result due to all sources of error and uncertainty in the modeling and simulation process; These assessments are collectively referred to as VVUQ."
- Oberkampf & Trucano, "Verification, validation, and predictive capability in computational engineering and physics,"
What we used
*Applied Mechanics Reviews* 57(5): 345–384 (2004) (asmedigitalcollection.asme.org, checked 2026-11-20), "Verification and validation (V&V) of computational simulations are the primary methods for building and quantifying this confidence. Briefly, verification is the assessment of the accuracy of the solution to a computational model. Validation is the assessment of the accuracy of a computational simulation by comparison with experimental data. In verification, the relationship of the simulation to the real world is not an issue. In validation, the relationship between computation and the real world, ie, experimental data, is the issue."
- Machine Design, "Standardizing Computational Models: Verification, Validation and Uncertainty Quantification" (machinedesign.com, checked 2026-11-20)
What we used
"The American Society of Mechanical Engineers (ASME) officially formed a committee to address the issue of verification and validation in 2001. The ASME V&V 10 Subcommittee in Computational Solid Mechanics undertook the task of writing proposed guidelines... By 2006, the first edition of ASME V&V 10, Guide for Verification and Validation in Computational Solid Mechanics, was published."; "The standards committee's name was changed in 2021 from V&V to VVUQ in order to highlight the importance of uncertainty quantification when analyzing the results of modeling and simulation activities."
- MatForge, "Validation and Verification for PDE Simulations: A Practical Framework" (matforge.org, checked 2026-11-20)
What we used
"Verification ensures your code solves the equations correctly (solving the equations right). Validation confirms your model accurately represents real-world physics (solving the right equations)."; "Verification must precede validation. You cannot validate an unverified code; doing so merely validates the buggy implementation."; "The Method of Manufactured Solutions is the gold standard for code verification in computational science."
- McKinsey, "Digital twins: the key to smart product development" (mckinsey.com, checked 2026-11-20)
What we used
"In advanced industries, survey data indicate that almost 75 percent of companies have already adopted digital-twin technologies that have achieved at least medium levels of complexity."; "Over the next five years, we estimate that about $30 trillion in corporate revenues will depend upon products that have not yet reached the market."; "In one survey, 75 percent of product development executives said that further digitization was a key priority for them."
- Interwork Software Solutions, "Digital Twin Manufacturing: Simulation ROI"
What we used
(interworksoftware.com, checked 2026-11-20), "McKinsey research documents digital twin programs cutting development time by up to 50%, reducing transportation and labor costs by up to 10%, and increasing revenue by as much as 10% for organizations building customer-facing digital twins (McKinsey)."; "NIST's Economics of Digital Twins research puts the total potential impact of digital twin adoption in U.S. manufacturing at $37.9 billion, against a documented $193.6 billion in combined maintenance costs and operational losses across the discrete manufacturing sector it examined (NIST)."; "Digital twins answer two additional questions that sensor monitoring alone cannot: what is causing this outcome, and what would happen if we changed this parameter?"; market size per market.us: the digital twin manufacturing market grew from $3.6 billion in 2024 and is projected to reach $42.6 billion by 2034, a 28.1% compound annual growth rate
- Vocal Media, "Digital Twin in Manufacturing Market" (vocal.media, checked 2026-11-20)
What we used
"This virtual factory mirrors more than 30 of BMW's global production sites in real time, and according to BMW Group itself, it has reduced production planning costs by up to 30%."; market size per Evolvance Market Research: the global digital twin in manufacturing market was valued at USD 8.12 billion in 2025 and is projected to reach USD 139.57 billion by 2035, a CAGR of 32.9%
- Signisys, "Digital Twins: Where the $34B Market Pays Off" (signisys.com, checked 2026-11-20)
What we used
"Fortune Business Insights values the global market at about $34 billion in 2026, growing at a 35.4% CAGR. Organizations using them report 65% less unplanned downtime, 62% better asset utilization, 90% faster decisions and 79% cost savings through predictive maintenance and real-time simulation."
- Cluster facts re-checked on disk (2026-11-20): #3 pid-tuning-simulation-first, the site's own rule: "control loops modelled and tuned in simulation before anything moves"
What we used
P alone leaves a steady-state error of a third of the step, forever; output = Kp × error + Ki × (error added up over time) + Kd × (how fast the error is changing); D on the measurement, not the error, to avoid derivative kick. #29 the-shape-of-the-move-trajectory-planning, S-curve commanded 1.058 s vs trapezoidal 0.857 s but settles at 1.058 s vs 3.042 s, 65.2% faster overall; residual 0.038° vs 0.734°, 19.2× smaller; for a 100 mm move at 100 mm/s with an acceleration of 500 mm/s², each ramp takes 0.2 seconds and covers 10 mm, leaving 80 mm of cruise, for a total of 1.2 seconds (Robotics Project Hub worked example). #10 robot-arm-control-sense-decide-move, UR5e-class six-axis cobot, 5 kg payload, 850 mm reach, 20.7 kg arm weight, pose repeatability ±0.03 mm per ISO 9283, force/torque sensor accuracy 4 N. #40 prototype-failure-post-mortem, four of ten hubs stopped transmitting between day 19 and day 23; the rail sagged under BLE transmit load, the sensor latched SDA low, the firmware waited forever; the failure was found in the field at 100× the validation cost (Semishops escalation: 1× / 5× / 20× / 50× / 100×+). #39 cobot-safety-risk-assessment, the chest limit 0.24 m/s at 20 kg effective mass, the 500 mm separation distance, the CE step. #6 power-design-battery-products, CR2032 230 mAh = 2,484 J. #17 sensor-calibration-drift, gyroscope measures rate, not angle; bias 0.1°/s adds 6°/min, 360° in 1 h