What a hybrid digital twin is, and when it is worth the extra work

A hybrid digital twin models the governing physics of an asset and uses data-driven components for the part the physics does not capture. Both halves are load-bearing: remove the physics and it is a statistical model, remove the learning and it is a simulation that slowly disagrees with reality.

It excludes two things that are often described this way. A statistical model with engineering rules bolted on top is not hybrid; the rules filter outputs rather than constrain the model. And a first-principles simulation that is manually re-tuned when it drifts is not hybrid either, because nothing is learning.

The word is worth claiming carefully because it is currently unowned. Siemens has established executable digital twin for its own architecture. Hybrid is used loosely enough that a buyer cannot tell from a proposal what is being offered, which is the same problem the wider term has and is described in digital twins for power plants.

The physics takes the parts that are known and stable: conservation of mass and energy, thermodynamic limits, established rate expressions, the geometry and connectivity of the plant. These are not learned because they do not need to be, and learning them from data would only produce a worse version of something already known exactly.

The data-driven components take what the physics genuinely does not describe. Fouling and degradation coefficients that drift. Parameters that are known in form but not in value for this specific machine. Local effects the model treats as uniform. Sensor bias. And the residual left when all of that is accounted for.

Drawing that line is the design work, and it is where hybrid models are done well or badly. A line drawn too far towards the physics produces a model that cannot adapt to the plant it is installed on. Too far towards the data and the constraints stop constraining anything.

The common pattern is straightforward. The physical model predicts what the plant should be doing under current conditions. The residual is the difference between that prediction and the measurement. Data-driven methods work on the residual rather than on the raw signal.

The advantage is that most of the variation caused by ordinary operation has already been removed. Load changes, ambient conditions and setpoint moves are explained by the physics rather than treated as anomalies, so what remains is closer to genuine unexplained behaviour.

That is also why this approach produces fewer false alarms on assets that vary constantly by design. A purely statistical model on a renewable-coupled plant spends its life flagging conditions it has not seen, which is technically correct and operationally useless. The mechanism is the same one described in reducing false positives in industrial anomaly detection.

A residual has more than one possible cause, and this is where honest implementations differ from confident ones. Time misalignment, an unaccounted stream, the wrong boundary or an incomplete model all produce residuals that look like process change. A useful system says which of those it can and cannot distinguish.

A fitted relationship holds where it was fitted. Outside that range it still returns an answer and carries no signal that it is now extrapolating, which is the property that makes purely statistical models dangerous on transient assets rather than merely inaccurate.

Physical constraints do not stop applying outside the training distribution. Mass still balances. Energy still balances. A conservation check that held at full load holds at twenty per cent, on a day the plant has never seen before, without anything having been fitted to that condition.

This is a bounded claim rather than a guarantee, and it is worth stating carefully because the field overstates it. Physics-informed models can also fail in extrapolation: the equations may be incomplete, the boundary conditions wrong, the parameters outside their valid range. What the constraints buy is that failures tend to be detectable, because an inconsistency shows up as a violated balance rather than as a confident wrong number.

The practical requirement that follows is out-of-domain detection and a defined fallback. A model that knows when it is outside its competence is more useful than one that is marginally more accurate inside it.

Mechanisms have to be modelled rather than inferred, and that takes domain engineering rather than software engineering. Somebody has to know what the governing relationships are for this equipment, which parameters matter, and what a physically implausible result looks like.

It is also slower to first output. A statistical model can be producing something from a historian export within days. A hybrid model needs the plant understood first, the tags mapped and validated, and the parameters calibrated against a reference measurement.

And it needs maintenance of a different kind. Not re-fitting when the distribution moves, which is the statistical model's recurring cost, but revision when the plant physically changes: equipment replaced, operating strategy altered, a new stream added.

Those costs are real and worth naming in a proposal rather than discovered in month three.

On an asset that runs a stable duty with a rich operating history, where the question is anomaly detection rather than attribution, and where nothing needs to hold outside the range already observed, a data-driven twin is cheaper, faster and sufficient. Choosing the more demanding architecture there is engineering vanity.

The case flips when any of three things are true. The asset spends significant time outside its historical envelope, which describes almost everything coupled to a renewable resource. The question is why rather than whether, because attribution needs a mechanism. Or the output has to be explained to somebody who was not in the room, whether an operator, an auditor or a counterparty in a dispute.

Green hydrogen and storage assets tend to satisfy all three at once, which is why the architecture argument arrives so quickly in that sector and so slowly in others.