manufacturing-digital-twins-shift-toward-autonomous-action-but-lack-governance-safeguards
Manufacturing Digital Twins Shift Toward Autonomous Action but Lack Governance Safeguards

Manufacturing Digital Twins Shift Toward Autonomous Action but Lack Governance Safeguards

A systematic review of manufacturing quality assurance research has found that factory digital twins are increasingly crossing the boundary from passive digital observation into physical machine intervention, yet none of the evaluated systems demonstrate a fully governed and auditable chain of assurance. The comprehensive investigation, published in the Journal of Engineering and Applied Science by researcher Mahrukh Muhaimeen Anindya, evaluated the operational mechanisms of digital twins and digital shadows designed for quality assurance across hundreds of published manufacturing studies. While engineers frequently deploy the term digital twin to describe virtual models that solely track process health or predict impending defects, contemporary applications are executing real-time interventions such as modifying toolpaths, altering welding feeds, halting fabrication processes, and commanding robotic sorting. However, the study reveals that the technological frontier of industrial computing is constrained not by whether connected software can intervene in the physical world, but by the critical absence of formal verification, auditability, operational boundaries, and authority governance required to trust autonomous decisions on the factory floor.

To rigorously investigate how digital models influence physical manufacturing, the review formulated the Embodiment-Authority-Assurance framework, establishing a structured methodology to differentiate among disparate forms of cyber-physical integration. Historically rooted in the foundational three-component construct introduced by Michael Grieves and John Vickers, digital twins have been formally differentiated from digital shadows based on the directionality of data communication. Whereas a digital shadow relies exclusively on physical-to-virtual data flows to mirror machinery behaviors, a true digital twin demands bidirectional interaction where virtual-to-physical pathways alter physical manufacturing. The newly introduced framework addresses how quality information is physically embodied into production, who or what holds decision-making authority over those interventions, and what empirical assurance establishes the safety, traceability, and credibility of the resulting action. Under this analytical model, physical embodiment traces mechanisms like parameter tuning and robotic manipulation, authority categorizes whether actions are manual, human-approved, or autonomously executed by algorithms, and assurance measures whether interventions are bounded, auditable, and verified against final product quality.

The author conducted a structured literature review spanning studies indexed in Scopus and Web of Science between 2017 and May 2026, identifying an initial corpus of 868 publications. Following the removal of duplicate records, 754 unique articles underwent title and abstract screening to filter out general production research that lacked substantive digital twin architectures, dedicated quality assurance mechanisms, or physical manufacturing interventions. Full-text screening of 419 candidate articles further isolated a final sample of 235 peer-reviewed studies containing sufficient operational detail to be coded across embodiment, authority, and assurance dimensions. To evaluate consistency, all 235 papers were subjected to a two-pass classification procedure using an ordinal five-level maturity scale, achieving approximately 94 percent agreement and a weighted Cohen kappa of approximately 0.93. Discrepancies were resolved by defaulting to the lowest maturity level explicitly supported by empirical mechanisms, workflows, or hardware pathways detailed within the published texts.

The framework categorizes industrial capabilities across five distinct maturity tiers: Level 1 digital shadows with physical-to-virtual monitoring without physical action; Level 2 advisory twins providing quality recommendations executed externally by personnel; Level 3 human-approved twins where virtual-to-physical actions require explicit operator sign-off; Level 4 bounded autonomous twins capable of executing local physical adjustments within predefined rules; and Level 5 governed adaptive twins possessing dynamic adaptability, explicit authority structures, complete audit trails, fail-safe fallbacks, and post-action verification. Across the 235 analyzed publications, the distribution was heavily concentrated in the foundational stages, with the majority remaining at Level 1 and Level 2 maturity. While a substantial minority achieved virtual-to-physical execution at Level 3 and Level 4, exactly zero studies among the 235 met all criteria necessary to reach Level 5 governed adaptive autonomy, revealing a stark technological plateau across modern manufacturing literature.

The concentration of studies within Level 1 digital shadows does not reflect primitive engineering, but rather highlights a paradigm where advanced computational modeling remains divorced from physical intervention. Many Level 1 systems leverage sophisticated sensing arrays, multi-physics finite element models, Kalman filtering, and machine learning to achieve remarkable observational fidelity without directly modifying machinery. Documented Level 1 implementations include sensory mapping across more than one thousand 3D concrete printing test samples, thermal field reconstructions during metal casting using inverse heat transfer calculations, and sparse in-line B-spline geometrical estimations in sheet-metal forming. Similar observational rigor characterizes laser powder bed fusion additive manufacturing, where digital shadows accurately estimate melt-pool geometry and predict internal porosity without altering laser power, recoater blade motions, or scanning velocities. These systems excel at defect localization, structural stress tracking, and tool wear diagnostics, yet their data streams remain strictly confined to physical-to-virtual pipelines.

Level 2 advisory digital twins advance beyond pure visualization by generating optimized recipes, corrective parameters, and maintenance schedules, yet they deliberately withhold virtual-to-physical commands until human engineers execute them manually. In discrete manufacturing, Level 2 platforms utilize genetic algorithms, neuro-fuzzy systems, and reinforcement learning to forecast surface roughness, mitigate drilling burr formation, or compute optimal laser energy densities for directed energy deposition. Comparable advisory architectures support continuous process industries, optimizing fluidized-bed drying, guiding biopharmaceutical fermentation, and modeling cooling parameters in industrial steel production. Despite formulating precise adjustments capable of preventing nonconforming components, these advisory architectures lack embodied machine authority. The analysis emphasizes that algorithmic recommendations and simulation-based control strategies do not constitute bidirectional maturity until the virtual model is granted an integrated physical mechanism to alter operations directly.

Bridging advisory algorithms and physical execution, Level 3 systems establish executable virtual-to-physical pathways that deliberately preserve mandatory human supervisory approval. In remanufacturing environments, Level 3 systems automatically analyze localized damage, generate robotic deposition toolpaths, and simulate repair sequences, yet require an operator to approve machine execution. Similarly, in high-throughput pipe extrusion, soft-sensor digital twins calculate screw-speed modifications in under 200 seconds—vastly outpacing the roughly 20-minute manual tuning intervals typical of expert operators—while still routing set-point updates through human confirmation to maintain operational accountability.

Level 4 systems demonstrate true autonomous physical execution, directly commanding machine controllers, industrial robots, and programmable logic controllers without real-time human intervention. Machining environments display extensive Level 4 maturity, with systems dynamically regulating computer numerical control feed rates, suppressing chatter, modifying spindle velocities based on in-situ vibration monitoring, and adjusting toolpaths via on-machine metrology to compensate for thin-walled plate deflection. In wire arc additive manufacturing and robotic welding, Level 4 twins dynamically correct heat inputs, wire feed rates, and travel speeds when weld pool dimensions drift beyond tight tolerances. Level 4 autonomy also encompasses automated material disposition and containment, where digital twins monitor extrusion anomalies to halt 3D printers mid-build, trigger pneumatic rejection valves to discard defective drilled parts, or command sorting systems to divert out-of-specification pharmaceutical tablets.

Despite the proliferation of Level 4 autonomous interventions, the review identified widespread assurance deficiencies that prevent these systems from achieving governed Level 5 status. Across autonomous studies, virtual-to-physical actions remain confined to localized, rigid rule sets that lack comprehensive system governance. Critical gaps include the widespread omission of immutable audit logs tracking what inference triggered an action, how parameters were modified, and whether the machine successfully executed the command. Furthermore, few autonomous architectures incorporate automated rollback policies to reverse parameter adjustments if unstable dynamics arise, or post-action verification routines to confirm that an intervention actually rectified a defect. While isolated investigations demonstrate notable features—such as automated coordinate reversal during aerospace disc machining or strict supervisory logging during tablet manufacturing—none unify traceability, dynamic boundaries, and independent quality verification into a cohesive governance framework.

The study clarifies that incorporating industrial robots or standardized digital thread architectures does not inherently elevate the maturity of a digital twin. Robotic manipulators represent an embodiment mechanism rather than an index of cognitive governance, appearing across all maturity levels from passive condition-monitoring targets at Level 1 to autonomous toolpaths at Level 4. Similarly, modern data interoperability frameworks—including OPC Unified Architecture, programmable logic controllers, manufacturing execution systems, and ISO 23247 standards—function as foundational infrastructure that facilitates physical connectivity but does not ensure decision credibility or intervention safety. The author notes that the study’s scope was bounded by selected search queries and indexed publications, and that findings reflect reported literature rather than undisclosed proprietary factory deployments. Ultimately, the synthesis demonstrates that establishing trustworthy digital twins requires shifting research priorities from building complex predictive simulations toward engineering transparent, auditable, and safely governed intervention frameworks.

Subject of Research: Systematic review of embodiment, authority, and assurance in manufacturing quality digital twins

Article Title: From digital shadows to assured intervention: a systematic review of embodiment, authority, and assurance in manufacturing quality digital twins

Article References: Anindya, M. M. (2026). From digital shadows to assured intervention: a systematic review of embodiment, authority, and assurance in manufacturing quality digital twins. Journal of Engineering and Applied Science, 73(1), Article 437. https://doi.org/10.1186/s44147-026-01295-x

Image Credits: AI Generated

DOI: 10.1186/s44147-026-01295-x

Keywords: digital twins, manufacturing quality assurance, cyber-physical systems, embodiment authority assurance, digital shadows, virtual to physical intervention, closed-loop control, additive manufacturing, computer numerical control, industrial automation, process control, systematic literature review, verification and validation, autonomous intervention, industrial robotics, traceability