Most nanobubble pilots produce a number and settle nothing. The equipment goes in, the block or the pond does better than expected, everyone agrees it worked, and by the time the capital request lands nobody can say whether the machine did it or the weather did. The result is unusable for a capital decision, which is the only decision the pilot was run to support.
The failure is almost never in the hardware. It is in the design: no control, or a control that was never comparable, or a measurement taken at the wrong place, or a window picked once the numbers were in. Those are fixable before anything is installed, and fixing them costs nothing. What follows is the design we would want applied to our own equipment, including the parts that make it harder for us to look good.
What makes a nanobubble pilot valid?
A pilot is valid when it can separate the treatment from everything else that changed. That requires a stated hypothesis, a matched control running at the same time, instrumentation that records the mechanism rather than only the outcome, and a run length set before the data comes in. Anything short of that produces an anecdote.
The hypothesis has to be specific enough to be wrong. "See whether nanobubbles help" cannot fail. "Dissolved oxygen at the emitter will hold above the untreated line through the irrigation set, and marketable yield in the treated strip will exceed the control strip by more than the strip-to-strip variation we already see" can fail, and that is what makes it worth running.
Measure the mechanism as well as the outcome, because the two answer different questions. If yield moves and dissolved oxygen did not, the mechanism was not oxygen and the result will not transfer to the next site. If dissolved oxygen moved and yield did not, the water was not your binding constraint — a real and useful finding, and one no vendor enjoys reporting. A pilot that only records the outcome cannot tell those cases apart, and a pilot that only records the mechanism has proved the injector works, which was never in doubt.
Write it down before the injector is energized.
- The hypothesis. Stated so it can fail.
- The measurement that decides it.
- The size of difference that would count.
- The end date.
Every one of those becomes negotiable the moment data starts arriving.
What is the right control?
The right control is the same crop, the same season, the same water source, and the same agronomy, split only by whether the water is treated. Last year's yield is not a control, and neither is the block next door on different soil. Where a split is impossible, alternate treated and untreated cycles instead.
This is the hardest part of the exercise, and where most pilots quietly fail. A historical baseline compares this year's weather, prices, variety, and management against last year's, then attributes the entire difference to a machine. A neighboring block compares two soils, two irrigation histories, and often two operators. Both feel like controls and neither is one.
What works is a split inside a single management unit: one field divided across the same soil series with the treated and untreated halves both under the same schedule, or one greenhouse bench set against another on the same bench system, or two ponds stocked from the same batch on the same feed table. Replicate the split if the site allows it — two treated and two untreated strips beat one of each by a wide margin, because a single pair cannot distinguish a treatment effect from ordinary spatial variation.
Where a physical split genuinely is not possible — a single reservoir, one recirculating system — alternate periods instead: treated, untreated, treated, untreated, long enough that a seasonal trend cannot masquerade as the pattern. It is weaker than a spatial control and far better than nothing. Say which one you used when you report the result.
What should I instrument, and how often?
Instrument the mechanism and the outcome separately. Log dissolved oxygen continuously at the point of demand on both arms, not at the injector outlet, because the daily minimum is what a spot reading misses. Record water temperature, flow, and applied volume alongside it, and sample the outcome on a fixed schedule set in advance.
Probe placement decides what the data means. A probe at the injector outlet measures the machine; a probe at the emitter, in the root zone, or at the pond bottom measures what the organism receives. Only the second is evidence about the crop. Why a concentration and a delivery rate are different quantities is worth reading before choosing where the probes go.
Log continuously rather than sampling. Demand swings with temperature, feeding, and the hours after an irrigation event, and a convenient mid-morning reading can miss the daily minimum entirely — the value that sets stocking density in a pond. Calibrate both arms' probes against the same standard on the same day, and swap the probes between arms partway through so a drifting sensor cannot be read as a treatment effect.
Verify the treatment itself as well as its effect. Ask for a bubble size distribution and particle count measured on your water, with the instrument, the date, and whether the sample was diluted stated on the report, plus a blank run on the untreated feed water so the count means something. Our classroom section on characterization and measurement techniques sets out what those instruments can and cannot resolve.
How long does a pilot need to run?
A pilot runs at least one full production cycle, because a partial cycle cannot show whether an early gain survives to harvest. Perennials and soils need longer, since root systems and microbial communities respond across seasons rather than weeks. Set the end date before starting, and publish it, so the stopping point cannot follow the data.
Short pilots systematically flatter the treatment. Early vigor is real and it is also the easiest thing in agriculture to observe and the least reliable predictor of what comes off the field.
A pilot that ends when the treated side looks better has measured the moment it looked better, not the season.
Different systems have different natural units: planting to harvest for an annual crop, stocking to grade-out for a recirculating system, a full range of loading and temperature for a treatment basin. Perennials and soil biology need more than one season, because root architecture and microbial communities respond on a timescale a single cycle does not reach.
Fix the end date in advance and hold it. A pilot that keeps running until the numbers turn favorable, or stops the week they do, has selected its own answer, and the selection is invisible in the final report.
What are the ways a pilot misleads?
Pilots mislead through novelty effects, unmatched controls, seasonal confounding, and windows chosen after the fact. A treated block also gets more attention, an untreated block on different soil was never comparable, a good season lifts both arms, and any long record contains a favorable stretch. Kairospace pilots are exposed to all four.
The novelty effect is the most under-rated. A treated block gets walked more often, inspected more closely, and managed more carefully. Some of that attention becomes agronomy, and the agronomy shows up in the yield attributed to the machine. The countermeasure is procedural: the same scouting schedule, the same interventions, the same people, on both arms, logged.
Seasonal confounding is the most common. A favorable year lifts treated and untreated together, and if only the treated arm was measured, the entire seasonal gain lands in the technology's column.
Window selection is the most tempting, and it applies to us. Any long dissolved-oxygen record contains a stretch where the treated line sits well above the untreated one; any season contains a fortnight where the treated block looks best. Choosing that stretch afterwards and presenting it as the result is not fabrication, and it is not evidence either. The defense is the pre-declared window.
One more, and it applies to us as much as anyone. A pilot whose only channel to the numbers is the vendor's own logger is a pilot whose result you cannot independently check. Own a probe, or have a third party own it.
What does a credible result look like?
A credible result states the control alongside the treatment, reports the mechanism and the outcome together, names the run length decided in advance, and shows the spread rather than a single average. It also reports what did not move. A result with no null findings in it has usually been filtered before publication.
The shape to look for is unglamorous.
- Both arms. Reported side by side.
- The dissolved oxygen record. Shown as a time series with its minima visible, not as a single average.
- Replicate-level numbers. Rather than one figure per arm, so a reader can see whether the difference is larger than the variation inside each arm.
- The confounder log. Attached, with a plain statement of what the pilot could not distinguish, because every pilot has one.
Apply that standard to any vendor's case study, and to ours. A result presented without its control, without the run length, and without a single measurement that failed to move has been through a filter — and a vendor who cannot produce the raw record when asked is asking to be trusted rather than checked.
None of this is expensive. Probes, a logbook, a split block, and a date written down in advance are the whole of it, and they are the difference between a number that survives a lender's questions and one that does not. Our pilot program page sets out how we scope and instrument one, and the design above is the one we would want a customer holding us to.