Can AI replace the jar test? What happened when we tried.

In 2015, I published a paper on using neural networks to predict coagulation performance. That work eventually taught me why AI cannot replace the physical jar test.

‍

Not in the way a chatbot invents a citation that does not exist. It was subtler and more dangerous for this application. The model produced answers that were mathematically defensible but chemically impossible: dose, pH, alkalinity, and removal combinations that no real water could produce together.

That was the first version of what eventually became Virtual Jar®. Once the “hallucinations” were discovered, we immediately went back to the drawing board and rebuilt it, so chemistry sits at the foundation, not machine learning.

I brought that story up during a live Q&A on August 20 with Brad Spangler, our VP of Sales, moderated by Chris Lacey. In the opening poll, a little over half the audience said they run physical jar tests every week. Their questions circled the same concern: what is a model doing with my physical jar-test data, and can I trust it?

A model is only trustworthy here if it is built on the chemistry. The jar test remains the ground truth; AI belongs beside it to interpret, extend, and pressure-test what the bench already tells us.

‍

WHY THE PHYSICAL JAR TEST STILL WINS

Jar testing's primary job is simple: guide the coagulant dose for the desired treatment performance. Most plants run the jar across a range of doses and set their dose from what they measure in terms of turbidity and organics removal performance.

The measured performance isn't all of it. You watch the jar — floc size, how fast it builds, how quickly it settles. A plant with a short flash mix needs a coagulant that reacts and settles fast; a plant with time to spare does not. That distinction shows up in the jar and not on a trend line.

Protocol discipline matters as much as the result. We dose neat with micropipettes rather than made-down stock solutions, because a diluted coagulant can begin hydrolyzing and reacting before it reaches the jar. That changes the active chemistry the test is supposed to measure.

‍

WHAT A MODEL CAN HONESTLY DO WITH THAT DATA

Here's the part fewer people have seen.

In 1999, Mark Edwards at Virginia Tech published a method for calibrating coagulation performance across coagulants — alum and ferric, the core commodity products. Five years later, Kastl and colleagues extended the approach to account for differences in TOC character. Both methods used error minimization, originally run in Excel's Solver, to fit chemistry-based parameters to measured removal data.

Then it sat. Very little changed after 2004, and none of it made its way back to the jar test an operator runs on their daily shift.

What we do now is run that method, with improvements, directly against physical jar test results. The calibration needs TOC or UV254 removal data because the model is estimating how much of the organic fraction is removable under the tested chemistry. It returns four parameters that describe that water-coagulant system.

Those four parameters are what the model simulates from. Nothing about your water is assumed — it's extracted from your coagulant and your water.

‍

WHY WE DON'T FIT THE MODEL TO PLANT HISTORY

An attendee asked directly what separates our approach from an AI plant simulation. This is where the 2015 story earns its keep.

A model fit purely to historical plant data learns correlations without constraints. It will extrapolate into conditions the chemistry forbids, and you won't know which solutions are reasonable. Something you can trust “most” of the time isn't useful when public health is the outcome.

So, the model does not calibrate against years of plant history. It starts from physical jar-test chemistry including the chemistry of the coagulant and then compares what the plant is measuring right now — turbidity and organics removal — against what the chemistry predicts for today's raw conditions. Add coagulant and you consume alkalinity; consume alkalinity and pH moves; move pH and the sorbable fraction of the TOC behaves differently. Those calculations run underneath every simulation, keeping the answers inside the range real water can produce.

One practical consequence: you don't need years of data to start.

‍

THE WORKFLOW THAT CHANGED HOW ONE PLANT DOSES

We've worked with a river-source utility for about three years. They already ran a physical jar test every Monday; the only change we made was that on Monday afternoon they also ran Virtual Jar® before setting the week.

What they found in their own data is the part worth repeating. Across four dates in a single year, with their coagulant dose staying within about a five-part range, TOC removal swung from around 60 percent down to 21 percent. Nothing was wrong with the coagulant or the operators. The water was simply harder to treat on some days.

The Organics Treatment Index is what provides insight during operations. It doesn't report TOC removal — it reports how much of the removable fraction they removed. On the low-removal days, it tells them they were already getting nearly all of what was available.

Without that, the instinct is to keep raising the dose to chase a number the water can't give you. This plant targets removal well above its regulatory requirement, because that helps them manage disinfectant residual and DBP levels. The index tells them the least-cost way to hit the target they choose — and when they've already hit it.

‍

LOOKING AHEAD

An attendee asked whether combining physical jar testing with simulations means fewer operator jobs. It does not, and the reason is deliberate: we do not build it to control chemical feed. Operators remain responsible for the dose decision. The tool exists to show them the chemistry behind the options, not to make the decision for them.

The analogy I keep coming back to is the road atlas. Getting from Chicago to St. Louis once took real effort just to plan the route. Now that's instant — but you still must drive, respond to conditions, and decide what to do when the map and the road do not quite match. The driving was always the job.

For anyone curious where to start: the barrier is usually organics measurement, not software. Coagulant demand, disinfection demand, and DBP precursors all trace back to organics, so UV254 or TOC capability is the real prerequisite. A benchtop UV254 unit runs a few thousand dollars, some plants already own one without realizing it, and the reading takes about two minutes a day.

Plenty is unsolved. Predicting filterability across coagulants is on the roadmap and not there yet — we're forecasting unit filter run volume with one client, four months in, and still separating what's chemistry from what's media age and backwash practice. Stay tuned.

If you're working out how much more your bench data and physical jar testing could be telling you, that is exactly the kind of problem our team works on every day. AI does not replace the jar test — it helps the jar test tell the whole story.