Using AI Metadata in Digital Twins

Why Semantics Matter When Using Generative AI

A digital twin model, on its own, is just names and types: RPM: double, Temperature: double, Friction: double. That’s enough for code to compile, but it tells an AI agent almost nothing. The agent can see that a number exists — it has no way to know what the number means, what a normal value looks like, or which properties actually matter for the question a user is asking.

Generative AI is good at reasoning over context it’s given, but it cannot reliably infer intent from bare identifiers alone. A property called P1 or even a reasonably-named one like Friction is ambiguous without semantics: friction measured how, in what units, and at what point does a value stop being normal? Attaching that context explicitly, through the attributes described in Annotating Digital Twin Models for AI, is what turns a schema into something an agent can actually reason about — closing the gap between “data exists” and “data is understood.”

What Good Metadata Looks Like

Good metadata is short, specific, and answers the questions an agent would otherwise have to guess at:

Metadata	Example	What it enables
Description	“Rotational speed of the turbine rotor.”	The agent knows what the property represents, not just its name and type.
Units	"RPM", "°C"	The agent reports and compares values correctly, and can reason about conversions or magnitude.
MinValue / MaxValue	0 to 20	The agent recognizes an out-of-range reading as anomalous, without being explicitly told what “too high” means.
Module Description	“Monitors rotational speed, temperature, and friction to detect unsafe operating conditions.”	The agent understands the model’s overall purpose, not just its individual fields.
KeyProperties	"RPM,Temperature,Friction"	The agent knows which properties to prioritize when a question doesn’t name one specifically.

What This Enables in Practice

With that context in place, an agent can move beyond simply reading back raw values:

• Answer natural-language questions correctly. “Is this turbine running hot?” requires knowing both the current Temperature value and its normal range — something only the metadata provides.

• Flag anomalies without hard-coded rules. MinValue and MaxValue let the agent recognize an unusual reading on its own, rather than a developer having to encode thresholds into every possible query.

• Prioritize what matters. In a model with dozens of properties, KeyProperties tells the agent where to focus when a user’s question is broad (“how’s this turbine doing?”).

• Communicate results clearly. Knowing the Units means the agent reports “1,850 RPM” instead of an unlabeled, ambiguous number.

In short, semantic metadata is what lets a generative AI agent treat your digital twins as genuinely understood data rather than an opaque set of fields — see Annotating Digital Twin Models for AI for how to add it to your own models.