Involving before Evolving: A Vision for Trustworthy Enterprise Digital Twin Engineering

• EDTs combine department- and enterprise-level models, goals, and autonomy as an acknowledged and directed System of Twinned Systems.The paper defines an EDT as analyzable and simulatable enterprise models that mimic real-world phenomena and support virtual what-if scenarios.Source: §1 · p. 1
• Michelin professionals navigate siloed data and processes while relying on tacit knowledge for what-if analysis and decisions.That knowledge can lack explainability and data grounding, and risks disappearing when individuals leave.Source: §1 · p. 1
• The fragmented information environment limits rapid scenario evaluation, uncertainty quantification, and robust long-term forecasting for informed decisions and resilience.Source: §1 · p. 1
• Stakeholders identified five needs: strategy support, financial what-if simulations, non-technical modeling, interoperability, and observability.These challenges were extracted through informal discussions with company stakeholders.Source: §1 · p. 1

• The socio-technical “involving before evolving” paradigm seeks organizational buy-in before advancing the EDT toward higher maturity and automation.The authors identify potential resistance in organizations that favor stability and established workflows.Source: §2 · p. 2
• Long-loop EDT predictions may inform decisions with severe, long-term consequences, raising the bar for stakeholder understanding and trust.Examples include restructuring a supply chain or opening a new market.Source: §2 · p. 2
• The first stage quickly gathers SME requirements and builds a low-maturity prototype with foundation models for early validation.The prototype integrates domain knowledge into simulation inputs and proposes LLMs, TSFMs, and uncertainty-aware Monte Carlo financial simulations.Source: §2 · p. 2
• After validation, the approach federates department-owned models through an ontological backbone to support semantic alignment and interoperability.The authors propose OML and related methodologies for federation.Source: §2 · p. 2
• The planned observability ecosystem makes data and predictions explainable, traceable, versioned, access controlled, consistent, and observable across silos.Proposed observability metrics and SME-oriented analysis tools, such as DTInsight, are intended to increase trustworthiness.Source: §2 · p. 2
• A decision-support prototype exists for marketing and finance; federation and broader department modeling are next steps toward trustworthy decision support.Source: §2 · p. 2

• The proposed minimum viable product captures informal requirements for early validation before investment in full verification and formalization.A tangible prototype is intended to involve stakeholders and their knowledge while the DT continues to evolve.Source: §3 · p. 2
• Foundation models can rapidly build interfaces and models or serve as forecasting and simulation placeholders where expert-developed models are unavailable.The paper proposes LLMs, TSFMs, and action-conditioned world models; it extends beyond using LLMs only as tool translators [20].Source: §3 · pp. 2–3
• The interface takes historical time-series data and offers a TSFM-predicted spline whose points and uncertainty ranges users can edit.Monte Carlo simulation samples the spline-defined ranges and propagates uncertainty through downstream expert-developed models to produce business-outcome distributions.Source: §3 · p. 3, Figure 2 · p. 3
• The functional prototype, rapidly built with agentic LLMs on company infrastructure, satisfies SME requirements for modeling scenarios from historical data with uncertainty control.Source: §3 · p. 3
• The prototype focuses on visual understanding for financial and marketing decisions over short and long time horizons.Source: §3 · p. 3
• The simple formalism helped secure buy-in and prompt cross-department discussion, refining requirements for multiple datasets and more precise TSFM predictions.The prototype is slated for extension to the ontological backbone.Source: §3 · p. 3

• Federation lets teams retain ownership of data, models, and services while committing to shared semantic contracts for interoperability.Source: §4 · p. 3
• Ontologies define component structure, typed interfaces for exchanged data, and domain rules that enable consistency checks and inference.The paper illustrates rules with finance approval requirements for a supply-chain demand forecast.Source: §4 · p. 3
• The federation must accommodate segregated departmental datasets and independently evolving systems through stable versioning and well-defined interfaces.These conditions support interoperability and reliability for cross-department what-if analysis.Source: §4 · p. 3
• Department-owned ontologies can interoperate through shared contracts and rules, with inference rules encoding semantics beyond taxonomy-based terminology alignment.Source: §4 · p. 3
• Ontology reasoning could flag degraded model output when historical inputs fail domain-specific fidelity requirements.The example contrasts a market needing ten years of history with one needing only six months.Source: §4 · p. 3
• OML is a DSL over OWL for vocabulary owners to define concepts, relationships, constraints, and basic rules with supporting tooling.The text says it reduces development friction while preserving semantic rigor for typed interfaces and automated checks.Source: §4 · pp. 3–4

5 Observability for Promoting Trust

• Trustworthiness is defined as stakeholders’ ability to inspect, understand, trace, and challenge EDT outputs while preserving data sovereignty and organizational accountability.This definition centers stakeholder oversight rather than treating trust as a property of prediction accuracy alone.Source: §5 · p. 4
• Monte Carlo simulation supports analysis such as sensitivity analysis, while curve editing lets stakeholders define inputs aligned with their domain knowledge.Prediction versioning records which tool version and datasets generated a final prediction.Source: §5 · p. 4
• Agentic LLMs and a containerized local time-series foundation model keep development and sensitive-data forecasting on company infrastructure.The paper presents local models as necessary for data sovereignty when processing critical business data.Source: §5 · p. 4
• The planned DTInsight extension combines lightweight OML modeling with a changing constellation visualization and reporting page to support lifecycle collaboration.The reporting page is intended to track EDT evolution and give stakeholders a basis for discussion while the system is built.Source: §5 · p. 4

Expected benefits

• The vision expects early prototype involvement to increase stakeholder buy-in, ontology-backed alignment to reduce errors, and federation to improve agility and departmental control.It also anticipates observable, traceable, privacy-respecting probabilistic simulations and greater enterprise resilience from data-driven decision support.Source: §6 · p. 4

Research agenda alignment

• The vision extends research on accessible Model-Based Engineering tooling for non-experts to organizational adoption and aligns with roadmap questions about collaboration and lifecycle engagement.The cited roadmap questions concern communication among DT engineering teams and keeping stakeholders informed and engaged.Source: §6 · p. 4

Related approaches

• Unlike an early SoS approach that reasons over needs to reduce design spaces before simulation and visual analytics, this work proposes prototypes for iterative exploration and adoption, then ontology-based federation.The comparison is with Franzén’s work.Source: §6 · p. 4
• The approach contrasts with Barat et al.’s agent-based ESL techniques by treating the EDT as a wrapper over organizational structure with simulation placeholders and SME-facing Monte Carlo interfaces.It emphasizes prototyping and trustworthiness concerns alongside SME validation.Source: §6 · p. 4

Potential future directions

• Action-conditioned world models are proposed as a possible future placeholder for simulating business behavior and exploring scenarios, although they are currently developed mostly on video data.Their usefulness for businesses is presented as an expectation, not a demonstrated result.Source: §6 · p. 4

7 Conclusion

• A working EDT prototype has stakeholder buy-in across Michelin departments and targets non-technical stakeholders divided among organizational silos.Source: §7 · p. 4
• The next step is to connect the prototype to an ontological backbone, incrementally formalizing requirements against the emerging ontology after prototype validation.This reverses some ontology-first approaches and is intended to help discussions converge toward shared formal language.Source: §7 · p. 4
• Open challenges include semantic heterogeneity and feasible consistency checks at scale, scenario provenance across departments, and evaluating outcomes across multiple deployments.The project cannot conduct the longitudinal multi-deployment study needed to assess whether EDT-informed decisions improve organizational outcomes.Source: §7 · p. 4