Ask whether AI beats physics for water problems and you will get an argument rather than an answer. The question is malformed. A physics based model and a machine learning model are not two ways of doing the same job, and the useful exercise is deciding which part of a problem belongs to each.
We hold a doctorate in each of the two disciplines, which is less a credential than an explanation of why this piece refuses to pick a side.
What a physics model is actually for
MODFLOW, HEC-RAS, EPANET and SWMM all solve equations that describe how water moves. That has three consequences worth stating plainly.
- It extrapolates. A calibrated groundwater model can be asked about a pumping rate nobody has ever tried, because the physics does not care whether it has seen that case.
- It explains. When the answer is uncomfortable you can trace it to a boundary condition or a conductivity value and argue about that, which is what a regulator or a funder will want.
- It is slow and hungry. It needs geometry, parameters and boundary conditions, and a transient run over a large domain is not something you do between meetings.
What machine learning is actually for
A learned model finds structure in a record. It does not know what water is, and that is the point.
- It interpolates well and cheaply. Filling a gap in a gauge record, downscaling a coarse product, estimating a value at an unmonitored point between monitored ones.
- It is fast once trained. Fast enough to sit inside an interface somebody uses while making a decision, which a full transient model usually is not.
- It fails quietly outside its training range. Ask it about a drought worse than any in its record and it will answer confidently and be wrong, with no signal that it has left familiar ground.
Where each one wins
Physics wins when the question is about a future that differs from the past. New abstraction, a proposed structure, a climate scenario, a policy that has not been tried. Those are the questions worth modelling and they are exactly the ones a learned model cannot answer.
Machine learning wins when the question is about the present or the near future under conditions that resemble the record. Forecasting the next few days, detecting an anomaly in telemetry, classifying land cover, filling a record so that something else can use it.
Where they work together
The useful configurations are hybrid rather than either.
- Learned models feed physical ones. Remote sensing and ML turn satellite imagery into the recharge and abstraction estimates a groundwater model needs, in places with no gauges.
- Learned models emulate physical ones. Train on a few thousand runs of a slow model and you get something fast enough for an interactive interface, with the physical model still available for the cases that matter.
- Physical models generate training data. Where observations are thin, simulation can produce the record a learned model needs.
The honest version of the AI claim
Most of what is sold as AI in water is a dashboard with a forecast in it. That is not a criticism. A good forecast in front of the right person is worth a great deal. But it is not a substitute for knowing how the aquifer behaves, and a vendor who will not tell you which of the two they are selling is worth another question.
The uncomfortable half of our position is the half worth paying for. Being told which parts of your problem AI should not touch is more useful than a longer list of the parts it might.
We build both layers in house, which is why we can put the boundary where the problem wants it rather than where our product line is. Our groundwater modelling and software and AI services pages set out each side, and where the output has to be a thing your team opens rather than a report they file, it becomes a water decision support system.