RQ-001
How much digital twin fidelity is actually required for predictive maintenance?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
Open Research Questions
Industrial Twin Lab publishes uncertainty instead of hiding it behind a product narrative. The catalogue is deliberately bounded to fifteen questions. Each question can accumulate hypotheses, experiments, negative results, replications, and revisions without being mistaken for a solved claim.
The 2026 frontier adds five questions about governed agent systems, reproducible context, simulation-system packaging, connected-product data contracts, and evidence for high-risk industrial AI. They respond to the Digital Twin System Framework (opens in a new tab), the Industrial AI Agent Manifesto (opens in a new tab), SSP 2.0 (opens in a new tab), and European data and AI obligations without treating any framework as a solved engineering method.
The 4 September 2026 review refreshes two evidence dependencies. Modelica 3.7 (opens in a new tab) is the language revision used for this research snapshot. Regulation (EU) 2026/1744 (opens in a new tab) amended the AI Act schedule: the relevant high-risk provisions apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I product systems. These dates are versioned research context, not legal advice or evidence that an industrial system is compliant.
RQ-002
When does physics-informed feature engineering outperform end-to-end deep learning?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-003
Can synthetic failures improve models without introducing dangerous simulation bias?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-004
Which features generalize across machines of the same class?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-005
How should confidence from simulation, historical evidence, and machine learning be combined?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-006
Can an AI Scientist autonomously design useful industrial experiments while remaining outside the control loop?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-007
How should digital-twin uncertainty propagate into AI recommendations?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-008
Can fleet learning work without centralizing industrial raw data?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-009
How can organizations distinguish correlation, causality, and physical mechanism?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-010
What evidence should be required before an industrial AI model is allowed into production?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-011
How can multi-agent orchestration remain subordinate to a shared safety hierarchy and named human authority?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-012
What operational context must be captured to replay and audit an industrial agent recommendation?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-013
Can SSP 2.0 and FMI 3.0 preserve enough simulation architecture and provenance for cross-tool replication?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-014
What machine-readable data and metadata contract is required for a connected product to support trustworthy twin evidence?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
RQ-015
What technical evidence should support risk management, human oversight, and traceability for high-risk industrial AI?
Status: conceptual. A useful answer must state its asset class, operating envelope, evidence sources, uncertainty, and conditions under which it fails.
Industrial Twin Lab maturity model
Level 0 — Connected Asset
Machine → Data. Data access is known; no inference claim is made.
Level 1 — Observable Asset
Machine → Data → Monitoring. Signals and quality limits are visible and traceable.
Level 2 — Digital Twin
Machine ↔ Digital Representation. The representation is verified and validated for a stated use.
Level 3 — Twin Lab
Twin → Simulation → Experiments. Experiments are isolated, reproducible, and limitation-aware.
Level 4 — Machine Intelligence
Twin + Experiment Fabric + AI Scientist + Fleet Learning. Recommendations survive independent validation and explicit engineering review.
Maturity is not a software purchase. P-101 does not advance because a model was trained; it advances only when evidence supports a defined engineering decision across relevant regimes. A higher level also inherits every lower-level obligation: data quality, traceability, simulation validity, uncertainty, and human authority.