Digital twins are commonly described as virtual representations of physical assets, production lines or entire factories. The usual proposition is compelling: connect real-time data, reproduce system behaviour, simulate alternative scenarios, predict future conditions and optimise operations before changing the physical process.
That vision is technically achievable.
Operationally, however, it is incomplete.
A digital twin does not become valuable merely because it reproduces geometry, physics, control behaviour or sensor signals with a high degree of accuracy. It creates industrial value only when it represents enough of the operational situation to improve a real decision.
That situation extends well beyond the physical asset. It includes the production plan, product variant, material condition, recipe, maintenance history, quality status, asset criticality, workforce availability, approved operating standards, current constraints and the consequences of acting—or failing to act.
Without this context, a digital twin may be an impressive engineering model while remaining a weak operational capability.
A Technically Accurate Model Can Still Support the Wrong Decision
Consider a digital twin monitoring a critical press in an automotive plant.
The model identifies a change in vibration behaviour and estimates that a component will continue to deteriorate over the next several production cycles. From a purely technical perspective, an immediate shutdown may appear to be the safest response.
The operational decision is more complex.
The responsible team must consider whether continued operation creates an unacceptable safety or quality risk. It must understand the priority of the current production order, the availability of qualified alternative equipment, the condition of downstream buffers and the consequences for customer delivery.
It must also determine whether the required component is available, whether appropriately qualified maintenance personnel can intervene, whether the work can be completed during a planned interruption and whether another failure elsewhere in the plant currently has greater operational criticality.
The physical model may understand the asset.
It does not automatically understand the decision environment.
This distinction is fundamental. Industrial operations are not governed by isolated technical truths. They are governed through managed trade-offs between safety, quality, production, maintenance, cost, delivery and long-term asset health.
A digital twin that ignores these relationships may generate technically sophisticated conclusions that cannot be executed responsibly.
Model Accuracy Is Not the Same as Decision Quality
Three different forms of validity should be distinguished.
Physical validity concerns whether the model represents the behaviour of the asset or process accurately.
Operational relevance concerns whether the model incorporates the conditions and constraints that materially affect the decision.
Decision legitimacy concerns whether the proposed response is compatible with approved procedures, risk limits, decision rights and escalation rules.
A model may be physically valid while being operationally incomplete. It may also generate an operationally reasonable recommendation that cannot be authorised under the organisation’s existing governance framework.
For this reason, the maturity of a digital twin should not be assessed only through prediction accuracy, simulation fidelity or visual sophistication. It should also be assessed through the quality, consistency and traceability of the decisions it supports.
Simulation Is One Layer of the Capability
Simulation remains one of the most important functions of a digital twin.
It allows engineers to test line configurations, evaluate cycle-time changes, investigate thermal behaviour, estimate wear, validate control logic and assess production scenarios without disturbing the physical process.
However, simulation answers a specific question:
What could happen under a defined set of assumptions?
Operational decision-making asks a broader question:
What should be done now, given the current state, constraints, responsibilities and risks?
Moving from the first question to the second requires operational context.
A simulation may show that increasing machine speed would raise output. The current operating context may reveal that the downstream quality station is already overloaded.
A degradation model may indicate that a maintenance intervention can be postponed. The production plan may show that tonight is the only feasible maintenance window during the next ten days.
A line twin may identify a sequence that minimises changeovers. Material shortages, customer priorities, tooling restrictions or qualification requirements may make that sequence impossible.
The mathematics may be correct.
The operational view may still be incomplete.
A Factory Is Not Only a Physical System
Many digital-twin programmes begin at the equipment layer. Sensors, PLC data, historian tags, engineering parameters and three-dimensional models are connected to a virtual representation of an asset or process.
This is a logical technical starting point, but it can encourage a misleading assumption: that the factory can be understood primarily through its physical behaviour.
A factory is a socio-technical system.
Machines operate within processes. Processes operate within production schedules. Schedules respond to customer demand and material availability. Maintenance work competes for labour, parts and time. Quality decisions depend on specifications, genealogy, process history and risk. Operators work according to standards but must also respond to abnormalities. Supervisors continuously balance immediate output against process stability and asset condition.
An operationally meaningful twin must therefore understand more than whether an asset is running.
It must know what the asset is producing, under which recipe and configuration, for which order, with which material and under which quality status.
It must distinguish between a genuine process abnormality and an approved temporary operating condition.
It must determine whether reduced line speed is caused by equipment deterioration, quality containment, material shortage, staffing constraints or an intentional production decision.
Without this semantic and operational layer, the data may be physically accurate but managerially ambiguous.
Operational Context Is Distributed Across Industrial Systems
No single industrial application contains all the information required to support a complex operational decision.
The PLC and SCADA layers may provide machine state, speed, temperature, pressure, alarms and control variables.
The historian may preserve time-series behaviour and process trajectories.
MES or MOM may provide the production order, product, routing, recipe, actual cycle, downtime classification, material genealogy and work execution status.
ERP may provide demand, production plans, customer priorities, material availability and cost information.
CMMS or EAM may provide asset hierarchy, maintenance history, open work orders, failure modes, spare-parts status and maintenance plans.
QMS or LIMS may provide specifications, inspection results, deviations, containment status and product-release information.
Workforce systems may provide availability, certification and qualification data. Engineering systems may contain design intent, configuration baselines and authorised equipment changes.
The objective is not to integrate every system merely because integration is technically possible.
The architecture should be derived from the decision that the digital twin is expected to improve.
For each decision, the organisation should identify:
- Which contextual variables materially change the outcome
- Which system is authoritative for each variable
- How frequently the information must be updated
- What level of data quality is required
- What should happen when information is missing, contradictory or obsolete
- Which assumptions may be inferred and which must be explicitly confirmed
A digital twin becomes operationally valuable when it connects enough of these perspectives to represent the decision environment with sufficient reliability.
Start With the Decision, Not the Model
A common starting question in digital-twin initiatives is:
What can we model?
A more disciplined question is:
Which recurring operational decision are we trying to improve?
The decision may concern whether an asset should continue operating under a degraded condition.
It may concern how production orders should be resequenced when material availability and capacity change.
It may concern whether a process deviation requires adjustment, containment, escalation or shutdown.
It may concern how a proposed line modification should be validated before launch.
It may concern how energy consumption should be balanced against throughput, product quality and delivery risk.
Once the decision is defined, the required model, context and governance become clearer.
The project team should determine:
- Who owns the decision
- Who provides the technical assessment
- Which information is currently used
- Which constraints must never be violated
- Which alternatives are operationally available
- How rapidly the decision must be made
- What uncertainty can be tolerated
- What evidence must be retained
- Which roles may approve, challenge or override the recommendation
Without this definition, a project may produce an advanced technical environment that operators and managers do not trust when operational pressure increases.
From Prediction to Maintenance Decision Support
Digital twins are frequently associated with predictive maintenance. The principle is sound: combine observed asset behaviour with statistical or physical models to estimate degradation and anticipate failure.
Prediction alone, however, does not determine the maintenance response.
Suppose a model estimates that a bearing has a high probability of failure within 120 operating hours. The estimate is relevant, but the required action depends on several additional considerations:
- The confidence and uncertainty of the estimate
- The safety and quality consequences of failure
- The criticality of the equipment
- The production plan during the relevant period
- Available redundancy or alternative routing
- Spare-parts availability
- Labour and specialist availability
- Intervention duration and complexity
- The next feasible maintenance window
- Approved risk limits and escalation criteria
The twin should therefore do more than present a remaining-useful-life estimate or failure probability. It should help decision-makers evaluate operationally viable alternatives.
These may include continuing under enhanced monitoring, reducing speed or load, modifying the production sequence, preparing parts and labour for a planned intervention, transferring work to another asset or stopping immediately because the approved risk limit has been exceeded.
This is the difference between an asset model and a decision-support capability.
The Twin Must Represent the Current Operating Reality
Trust in a digital twin does not arise from model accuracy alone. It also depends on whether the model reflects the current operating state.
A simulation can remain mathematically valid while becoming operationally obsolete because the equipment configuration has changed, the tooling has been modified, the production routing has evolved, the recipe is outdated or a temporary engineering change was never incorporated.
This is why master data, configuration control and lifecycle governance are not secondary administrative concerns. They are part of the technical integrity of the twin.
The organisation must be able to answer:
- Which asset configuration does the twin represent?
- Does it reflect the current tooling and control logic?
- Are recipes and routings aligned with actual shopfloor practice?
- Are sensor mappings and calculation rules still valid?
- Who approves changes to the model?
- How are temporary modifications represented?
- How is model performance monitored?
- What happens when the physical system and the digital representation diverge?
- Under which conditions must the model be recalibrated, restricted or retired?
Without disciplined lifecycle management, the twin gradually becomes a digital memory of how the operation used to function.
The more sophisticated the model appears, the greater the risk that users will treat obsolete conclusions as authoritative.
Abnormality Requires a Standard and a Context
Many factories collect enough data to describe normal behaviour. The more difficult task is interpreting abnormal conditions correctly.
A machine operating below standard speed may indicate deterioration. It may also be running under an approved temporary mode.
A temperature deviation may represent a process risk. It may also be expected for a particular material batch, recipe or product variant.
A line stoppage may indicate equipment failure, quality containment, planned tool change, material shortage or an intentional safety intervention.
A twin cannot classify these situations reliably from physical signals alone.
This is where Lean, MES/MOM and digital-twin thinking should converge.
Lean provides the logic of standards, abnormality, flow, escalation and structured problem-solving.
MES/MOM provides execution context: what should be happening, what is actually happening, which order and material are involved, and why the operation has deviated.
The digital twin evaluates physical consequences and alternative scenarios within that context.
Governance determines who may act, under which conditions and with what level of traceability.
Without a clear standard, abnormality is difficult to define.
Without reliable execution data, the actual operating state is difficult to interpret.
Without decision ownership, the recommendation has no legitimate path to action.
The Pilot Is Not the Industrial Capability
Digital-twin pilots often perform well because their scope is controlled.
They may cover one asset, one product family or one engineering scenario. Data is manually cleaned. Subject-matter experts remain close to the project. Assumptions are understood. Model changes are managed informally.
The challenge emerges when the capability must scale.
New products are introduced. Equipment configurations evolve. Master data differs between sites. Local operating practices are not standardised. Models require continuous calibration. Ownership moves from the project team to operations. Recommendations begin to influence real production decisions.
At this stage, the principal challenge is no longer simulation.
It is architecture, data ownership, configuration management, process integration, accountability and operational adoption.
A successful pilot proves that a model can produce useful results under controlled conditions.
Industrialisation must prove that the organisation can operate, maintain, validate, challenge and trust the capability over time.
From Digital Replica to Governed Decision System
A mature digital twin should not be treated as an isolated visualisation, engineering application or analytics model.
It should form part of a governed operational decision system.
Such a system normally requires:
- A reliable representation of the physical asset or process
- A current view of the operational state
- Production, maintenance, quality and business context
- Explicit assumptions, constraints and decision logic
- Defined ownership for reviewing and acting on recommendations
- Traceability of model versions, inputs, outputs and final actions
- Feedback on outcomes to support validation and continuous improvement
This does not require full autonomy.
In many industrial environments, the most valuable digital twin will not be the one that makes the decision independently. It will be the one that enables a responsible human decision-maker to evaluate alternatives more rapidly, consistently and transparently.
The objective is not to eliminate professional judgement.
It is to improve the evidence, discipline and traceability with which judgement is exercised.
AI Increases the Need for Operational Discipline
As artificial intelligence becomes more closely integrated with digital twins, the need for context and governance becomes even greater.
AI may identify patterns, estimate risks, detect anomalies, generate scenarios or propose interventions. However, its output still enters an industrial environment governed by procedures, responsibilities, technical limits and real consequences.
An AI-enabled twin should therefore distinguish clearly between:
- Observed facts
- Inferred conditions
- Predicted outcomes
- Simulated scenarios
- Recommended actions
- Approved operating constraints
It should make uncertainty visible rather than presenting estimates as certainties.
It should not recommend actions outside validated operating procedures without appropriate escalation.
It should preserve accountable human authority for safety-critical, regulated or high-impact decisions.
It should also record the model version, relevant inputs, assumptions and reasoning path that produced a recommendation.
The advantage will not come from creating the most visually impressive twin or applying the most sophisticated algorithm.
It will come from governing how models influence operational decisions.
A Practical Readiness Test
Before industrialising a digital-twin capability, an organisation should be able to demonstrate that:
- The operational decision to be improved is clearly defined.
- The decision owner and process owner are identified.
- The necessary production, maintenance, quality and asset context is available.
- Relevant master data reflects current shopfloor reality.
- The model has a permanent owner beyond the original project team.
- Assumptions, uncertainty and operating limits can be explained.
- Recommendations can enter existing operational workflows.
- Decision rights, overrides and escalation rules are explicit.
- Actions and outcomes can be recorded for validation and learning.
- Model changes, recalibration and retirement are governed.
- The value case is connected to improved decisions rather than better visualisation alone.
The Real Twin Is Operational Understanding
A digital twin should represent more than the behaviour of equipment.
It should represent enough of the operational situation to support a responsible, executable and traceable decision.
This does not require a perfect virtual reproduction of the entire factory. That ambition can become prohibitively complex, slow to maintain and impossible to govern.
It requires deliberate selection of context.
A useful twin understands what is happening, what should be happening, which constraints matter, which alternatives are feasible, what uncertainty remains and who owns the final decision.
Simulation remains essential. Prediction remains valuable. Artificial intelligence can extend both.
But simulation without operational context is only a sophisticated engine of possibilities.
Industrial value appears when those possibilities can be translated into governed action under real shopfloor conditions.
The decisive question is therefore not how accurately the twin represents the asset.
It is whether the organisation can rely on it to make a better decision when safety, quality, production, maintenance and delivery priorities are in conflict.
#DigitalTwin #OperationalExcellence #SmartFactory #IndustrialMaintenance #MES #ManufacturingExcellence #AssetManagement #Reliability #IndustrialAI #LeanManufacturing