Industrial AI Safety Architecture

Industrial Twin Lab separates observation, experimentation, validation, and operational authority. Information may move from a physical asset into a controlled digital environment; an experimental result does not travel back as a control command. The full reasoning path is longer than a dashboard pipeline:

Physical Asset → Digital Twin → Isolated Experiment Lab → AI Scientist → Evidence → Human Decision → Validated Deployment

Industrial Twin Lab safety boundary

Diagram notation

  • Solid line = directional flow
  • Dashed boundary = isolated experiment zone
  • Signal-orange stop marker = human authority / safety boundary
  1. OT Control Zone

    P-101 is a fictional demonstration asset. PLC, DCS, and SCADA retain deterministic plant control under existing authority.

    Safety boundary

  2. Data Access Zone

    Sensors, gateway, read-only historian, OPC UA, and MQTT provide controlled acquisition and quality context.

  3. Twin Zone

    Twin Capsule, digital twin, and simulation runtime represent behaviour within a declared fidelity and envelope.

  4. AI Experiment Zone

    Isolated Twin Lab services include local models, feature engineering, model training, experiment tracking, and AI Scientist. They have no control authority.

    Isolated experiment zone

  5. Validation Gate

    Human in command. Engineering review, approval, and a signed model determine fitness for a stated inference use.

    Human validation gate

  6. Inference Zone

    An approved local model runs only as the reviewed version within its approved boundary, with no path back to control.

Experimentation is isolated from operational control. There is no direct control path from AI or an experiment result to the physical machine or OT Control.

Human engineering review remains in command of every physical-machine decision.

Six ordered zones keep research inquiry outside the control loop. Evidence may move forward for review; no experimental or inference result has a return path to P-101 or OT Control.
P-101Fictional demonstration asset

Real-world to digital-world evidence flow

  1. Physical Asset
  2. Digital Twin
  3. Isolated Experiment Lab
  4. AI Scientist
  5. Evidence
  6. Human Decision
  7. Validated Deployment
A one-way real-world to digital-world path: controlled data enters the lab, while evidence reaches a human decision before any separately validated deployment.

Zoned boundary

ZoneBoundary roleRepresentative elementsPermitted role
OT Control ZoneSafety boundaryP-101 fictional demonstration asset, PLC, DCS, SCADADeterministic plant control under existing authority.
Data Access ZoneControlled flowSensors, Gateway, Read-only Historian, OPC UA, MQTTControlled, read-oriented acquisition and quality context.
Twin ZoneControlled flowTwin Capsule, Digital Twin, Twin Runtime, Simulation RuntimeRepresent behavior within a declared fidelity and envelope.
AI Experiment ZoneIsolated experiment boundaryLocal Models, Feature Engineering, Model Training, Experiment Tracking, AI ScientistGenerate and test hypotheses without control authority.
Validation GateHuman authority gateEngineering Review, Approval, Signed ModelDecide whether evidence is fit for a stated inference use. Human in command.
Inference ZoneApproved inference boundaryApproved Local ModelRun only the reviewed version within its approved boundary, with no path back to control.

The separation is both logical and architectural. Access credentials, networks, software lifecycles, data paths, and approval records should make it difficult to confuse a research experiment with a production control change. Cloud resources may extend the lab, but must not become an undeclared dependency for critical local data, knowledge, experiments, or inference.

Training an AI model and granting an AI system authority over machinery are completely different activities.

For the fictional P-101 teaching asset, suction pressure, discharge pressure, flow, power, temperature, speed, and vibration can be copied into a bounded Twin Capsule. The twin may replay history or simulate a hypothesis about bearing degradation. The AI Scientist may request experiments and compare Evidence Packages. None of these actions changes a setpoint, starts a pump, bypasses an interlock, or certifies safety. An identified engineer decides whether further testing, maintenance inspection, or a separately validated inference deployment is justified.

Evidence crosses the gate, not authority

A candidate crosses the Validation Gate only with a reproducible record: asset and twin versions, dataset and feature versions, experiment configuration, validation strategy, performance by operating regime, uncertainty, limitations, explainability, and provenance. Approval is scoped. Evidence for detecting P-101 bearing degradation under a known envelope does not authorize control, another failure mode, another machine, or an unseen regime.

See the Twin Capsule, Experiment Fabric, and AI Scientist for the records on each side of this boundary.

Research direction / not deployed

Digital Triplet: research direction only

Physical Machine + Digital Twin + AI Scientist describes a research direction, not a deployed autonomous capability. It grants no control authority and remains subject to the same validation boundary and human engineering decision.