What a physics constraint does that another input cannot.

Black-box industrial AI usually starts from the assumption that enough historical data will eventually reveal the right pattern. That works in some domains, but electrolyzers operate inside hard physical constraints. If the model does not understand those constraints, it can learn correlations that fail exactly when the plant enters a new regime.

That is why teams often see one of two failures: the model misses the early warning signal entirely, or it predicts behaviour that engineers immediately distrust because it violates what the system can physically do.

Electrolyzer performance is shaped by kinetics, transport, temperature, pressure, and balance-of-plant interactions. A practical analytics stack should know that current density, gas purity, voltage drift, and thermal behaviour are not independent features. They are coupled.

Physics-first models give that coupling a structure. Machine learning can still add value by correcting model mismatch, learning plant-specific behaviour, and ranking abnormal scenarios, but it operates inside a physically coherent frame.

A statistical model given current density, temperature, pressure and voltage as four inputs will find the relationships between them that its training history contains. Those relationships are real, and they hold where the history holds. What the model does not have is the reason they exist, so it has no way to know which combinations are impossible rather than merely unseen.

A conservation check behaves differently. Mass balances. Energy balances. Those hold at twenty per cent load on a day the plant has never run, without anything having been fitted to that condition, and a set of readings that cannot be reconciled against them is inconsistent whether or not the pattern is familiar.

This is a bounded claim and the field routinely overstates it. A physical model can be incomplete, its boundary conditions can be wrong and its parameters can sit outside their valid range, and it can then be confidently wrong in its own way. What the constraints buy is a different failure mode: an inconsistency tends to surface as a violated balance rather than as a plausible number, and a violated balance is something an engineer can investigate.

A physics layer costs domain engineering, a longer path to first output, and revision whenever the plant physically changes. On an asset that runs a repeatable duty with years of history and instrumented failures behind it, that expense buys little a fitted model would not have found, and saying so is part of being trusted on the cases where it does matter.

It is also the wrong tool where the physics is not actually known. Modelling a mechanism that has never been characterised produces a confident guess with equations attached, which is harder to argue with than a statistical output and no more correct. Where the mechanism is unclear, the honest structure is a physical model of what is understood, data-driven correction of what is not, and an explicit statement of which is which.

The claim worth making is narrower than the one usually made. Physical constraints reduce implausible outputs and make some inconsistencies visible. They do not guarantee correct behaviour outside the envelope a model has been validated for, and results from outside it should be reported as uncertain rather than assumed to hold.

Operators do not need another dashboard that says a score changed from 0.73 to 0.81. They need to know what is changing, why it matters, and what action is worth taking next.

That is where physics-first AI earns trust. It turns analytics from pattern recognition into engineering guidance, which is the threshold most plant teams need before they will rely on a system in daily operations.

When the physics and the measurements disagree, the disagreement itself is the useful output. It says the picture is inconsistent. It does not say which part is wrong, and treating it as though it convicts the instrument is a shortcut that produces confident misdiagnosis.

The candidates are worth naming: a faulty reading, a stream that is not metered at all, a boundary drawn in the wrong place, clocks that disagree between two systems, or physics the model does not carry. Each has a different owner and a different remedy, and narrowing to one of five is a substantial improvement on an alert that says a score moved.

That is the practical difference for an operating team. A statistical anomaly asks someone to go and look. A residual against a physical model asks a shorter question, arrives with the evidence that produced it, and can be explained afterwards to an engineer, an auditor or a lender in terms that do not require them to trust the model.