When Physics-Informed Models Fail

Your supervisor ends the mission with a warning, not a victory lap. 'Physics-informed models are only as good as the physics you chose. If we'd assumed a single pumping well, but there's actually a second, unmapped well nearby pulling water the other way, our constraint would be actively wrong, and we'd trust it more, not less, because it looks so principled.'

The biggest risk in physics-informed modelling isn't too little constraint, it's the wrong constraint, applied with false confidence. If your governing equation doesn't actually match the real system (missing wells, unknown geological boundaries, seasonal recharge you didn't account for), the model will still produce smooth, physically 'consistent' answers, just consistent with the wrong physics.

It is like confidently giving someone perfect directions using an old map that is missing a brand-new road. Every instruction sounds authoritative and follows the map's rules exactly, but it is quietly wrong the whole way, because the map itself, not the reasoning, was flawed.

This is a known failure mode across physics-informed machine learning: the technique reduces one kind of error (physically implausible predictions) but can mask a different one (systematically wrong predictions from a misspecified physical model). Careful researchers always validate their governing assumptions against independent evidence, not just against the same data used to fit the model.

Explain to a fellow researcher why 'the model respects Darcy's Law' is not the same guarantee as 'the model is correct'.