Editorial: Digital twins: from predictive to prescriptive an editorial on closing the decision gap in maintenance

What the paper says

The benefit of a digital twin is more than a live 3-D model on a control-room screen. Yet in many plants today, that is exactly where the technology stops. Operators watch color-coded heat maps and vibration trends. Analysts export comma-separated files for offline study, and maintenance planners still build their weekly schedules in spreadsheets. Data are available; however, timely, defensible action is lacking.The main challenge facing many maintenance organizations as they move into the next phase of digital-twin maturity is this disconnect between sensing and doing. The “predictive” problem has been largely solved by the industry. Algorithms can flag that a bearing is degrading or that fouling is developing in a heat exchanger. The “prescriptive” problem, where prediction is translated into a specific, risk-ranked work recommendation that a planner can approve, a technician can execute and an auditor can trace back and justify, is not yet resolved. This editorial argues that the real value of a maintenance-focused digital twin lies in closing precisely this gap and outlines the governance and organizational conditions that are crucial for success.Predictive maintenance uses condition data and machine-learning models to estimate when a failure is likely to happen. The output is typically a remaining-useful-life estimate or a probability-of-failure estimate. This is genuinely useful, but it answers only one question: “What will fail, and roughly when?” It does not answer the questions that maintenance leaders must resolve every day: What should we do about it? When exactly should we intervene? What are the cost and safety trade-offs of deferring versus acting now? Which crew, which spare and which time window?Prescriptive maintenance answers these follow-on questions. A prescriptive digital twin processes the same sensor streams, historical data and failure-mode libraries that feed a predictive model. However, it adds optimization and simulation routines that weigh risk, cost, resource availability and production schedules to generate a ranked set of recommended actions. Critically, each recommendation carries an auditable justification. The twin should explain why it suggests replacing a coupling next Tuesday rather than running to the next planned shutdown. Van Dinter et al. (2022) define prescriptive maintenance as “predictive maintenance with a module to prescribe an action plan,” noting that while the potential impact on service, cost and safety is optimal, the method is also the most complex to implement. That complexity is not primarily algorithmic; it is organizational and governance-related.A prescriptive twin is not a single model but a layered system. At the base sits a real-time data-integration layer that combines sensor telemetry, SCADA signals, laboratory results and maintenance work-order history into a unified asset state. Above this, physics-based or hybrid models estimate degradation trajectories and remaining useful life. The prescriptive layer then applies decision logic, often formulated as constrained optimization, simulation or reinforcement-learning agents. It then evaluates alternative maintenance actions against an objective function balancing reliability, cost and safety (Abd Wahab et al., 2024).Padovano et al. (2018) demonstrated an early prototype of this architecture in an oil-and-gas turbomachinery plant, where a digital-twin-based decision-support system combined real-time sensor preprocessing, predictive analytics, scenario simulation through the twin and a recommender module that proposed optimal maintenance and production schedules. More recent implementations integrate directly with enterprise asset management and planning systems so that recommendations flow into executable work orders rather than simply sitting in a dashboard.Closing the loop is what distinguishes a prescriptive twin from a predictive one. When a recommended action is executed, the outcome feeds back into the twin, updating its models and improving future recommendations (van Dinter et al., 2022). This continuous learning cycle is essential; without it, the twin's advice drifts from operational reality.No prescriptive system should be deployed without a governance framework that addresses transparency, accountability, validation and alignment with regulations. This is the most underestimated area, and it is where twin programs most frequently lose organizational trust.Transparency and explainability. Every recommendation on the twin issues must be traceable to its inputs, models and decision rules. A planner who cannot understand why the system is recommending an early bearing replacement will ignore it. Explainability is not a luxury feature; it is a prerequisite for adoption. The EU's Artificial Intelligence Act (Regulation 2024/1689) now formalizes this expectation for high-risk AI systems. A prescriptive twin that influences safety-critical maintenance decisions in sectors such as energy, transport or chemicals will increasingly fall within its scope. A recent compliance study applying six European digital laws to twin deployments found “persistent gaps in the transparency, explainability and accountability of AI-driven components,” even among organizations that had invested heavily in the technology (Jørgensen and Ma, 2025).Accountability. When a twin recommends deferring maintenance and the asset subsequently fails, who is responsible? The algorithm developer? The reliability engineer who approved the recommendation? The operations manager who accepted the risk? Current legal scholarship highlights that digital-twin technology “has accelerated ahead of a parallel regulatory framework, leaving uncertainty regarding accountability, privacy, data management and responsibility” (Ahmed and Uddin, 2025). Organizations must define accountability chains before deployment, not after a failure investigation. A practical approach is to treat the twin's output as a recommendation, never an instruction, and to require a human sign-off that is recorded alongside the twin's rationale. This “human-in-the-loop” pattern preserves professional judgment while still benefiting from the twin's analytical capabilities.Model validation and lifecycle management. A prescriptive twin is only as trustworthy as its models. Physics-based models require periodic recalibration as operating conditions change; data-driven models are susceptible to concept drift as equipment ages or process regimes shift. Governance must require model-validation periodic reviews and define clear criteria for retraining or retiring a model. ISO/IEC 30173:2023 now provides a shared vocabulary for digital twins, and the forthcoming ISO/IEC 30188 reference architecture will offer further structural guidance (ISO/IEC, 2023).Regulatory alignment and audit readiness. In regulated industries such as power generation, pharmaceuticals and aviation, maintenance decisions are subject to external scrutiny. A twin that generates recommendations must produce records that satisfy the same documentary standards as a human-authored maintenance plan. This includes versioned model configurations, timestamped input data and archived decision logs. Building audit readiness into the twin's architecture from the outset is far cheaper than fixing it after a compliance gap is discovered.Organizational readiness. A study integrating academic literature with in-depth practitioner interviews concluded that “organizational barriers outweigh technical challenges” in the adoption of AI-enhanced digital twins for maintenance (Chen et al., 2025). The most frequently cited obstacles were not algorithmic sophistication or sensor coverage but change management, unclear ownership and a shortage of personnel who can bridge reliability engineering and data science. Without deliberate investment in cross-functional teams and role clarity, twin programs stall after a successful pilot.Data quality and integration. The prescriptive layer is extremely sensitive to data quality. Sensor drift, inconsistent failure coding in the CMMS and missing contextual data such as which operating mode the asset was in when a reading was taken can turn a well-designed model into a source of misleading advice. Integration across OT and IT systems remains painful; proprietary protocols, legacy historians, and inconsistent naming conventions create friction at every boundary. Systematic reviews consistently identify data variety, asset complexity and the absence of standardized data protocols as top-tier challenges (van Dinter et al., 2022; Abd Wahab et al., 2024).The human-in-the-loop imperative. Trust is earned, not configured. Maintenance technicians and planners bring decades of tacit knowledge about how equipment actually behaves in their specific plant. A twin that overrides or marginalizes this expertise will be resisted, regardless of its theoretical accuracy. Successful deployments treat the twin as a decision-support tool, a highly capable advisor whose recommendations are reviewed, questioned and occasionally overruled by experienced professionals. Workforce upskilling is essential: planners need enough statistical literacy to interrogate a confidence interval, and data scientists need enough process knowledge to recognize when a model is extrapolating beyond its training domain (Chen et al., 2025).What a mature twin cannot do. Even a well-governed, well-integrated prescriptive twin has boundaries. It cannot reliably predict failure modes it has never seen. It cannot compensate for sensors that are not installed. It cannot resolve conflicting business priorities, whether to favor production throughput or equipment longevity, without human judgment. Setting honest expectations at the outset protects the program from the credibility collapse that follows over-promise. The global digital-twin market may be growing at nearly 40% annually (Hexagon, 2025), but market enthusiasm is not a substitute for careful implementation.The shift from predictive to prescriptive is not a technology upgrade; it is an organizational transformation. It requires that maintenance leaders invest as heavily in governance, data quality and workforce capability as they do in algorithms and visualization platforms. Three practical steps can accelerate the journey. First, start with a single critical-asset class such as rotating equipment in a refinery, for example, and build the full prescriptive loop end to end before scaling horizontally. Second, establish a governance charter that defines model-validation cycles, accountability for recommendations and audit-trail requirements at the outset. Third, co-design the twin with the people who will use it: planners, technicians and reliability engineers whose daily judgment the system must complement, not replace.Finally, a note on where research itself must catch up. The academic literature on digital twins is overwhelmingly technical, focused on algorithms, architectures and sensor integration. Far less attention has been paid to the managerial, governance and organizational dimensions that, as this editorial has argued, ultimately determine whether a twin delivers value or fails. We need rigorous empirical work on how accountability frameworks function in practice, how maintenance organizations absorb prescriptive tools without marginalizing professional expertise and what governance structures prove durable beyond the pilot phase. Until the research agenda broadens to match the sociotechnical reality of twin deployment, the gap between laboratory promise and plant-floor impact will persist.LLMs are used for English editing and searching for some suitable references