Harmful algal bloom early warning · desalination intakes

The bloom is already in the water.
Know days before it reaches the intake.

BloomShield is a response-timing and production-loss decision layer for seawater desalination. It reads the water the plant already measures, learns that site's own seasonal rhythm, and turns a bloom risk into one operational answer: act now, or wait.

It does not forecast the bloom, does not sell hardware, and does not treat your water. The plant keeps every lever; we make the timing call.

Open the live plant panelSee what we measured

Every number below is either measured on a blind holdout of real records, or a documented negative. Nothing is projected, and nothing is priced in this page.

How it works

Four steps, no new physics. The plant's existing record is the input; the output is a timing decision a shift supervisor can act on.

  1. 01

    Read the water

    The plant's own intake record is the input. Nothing is bought to start — we begin with the history the plant already logs.

  2. 02

    Learn its rhythm

    Every site has a seasonal baseline. We learn the baseline and flag when the water departs from it in the direction that precedes a bloom.

  3. 03

    Decide, don't just alert

    A cost–loss rule weighs an avoided production loss against the price of a false alarm, and returns one answer: act now, or wait.

  4. 04

    Record it

    What was seen, what was recommended, what happened. The audit trail finance and regulators ask for after an event.

intake record→seasonal baseline→risk→decision→operator

What we reached

Measured on 13 years of public real-world records, with the final test period held out completely — no tuning against it, no second look.

0.887AUCBlind holdout — 572 days the model never saw during training.
69.4%precision at 55.8% recallSame holdout, at the operating point the decision layer uses.
7–12 daysfirst warningAhead of the event on 6 of 8 major blooms in the untouched test period.
68bloom episodes catalogued2013–2026 across four public shore stations — 51 of them major. This coast does this every year.

And the parts that did not work

  • Satellite alone: replicated on Gulf water and scored no better than chance. Ocean colour is context and tracking — it never feeds the onset model.
  • Crossing to a new site is not free: without local calibration, transfer precision drops to about 58%.
  • Untested on Saudi water. We have zero measurements there, and we say so rather than extrapolate.

We publish our negatives because a warning system is only worth trusting if its failures are on the record too.

What it runs on

Free and public data, plus the plant's own records. No purchase is required to run the version that produced the numbers above.

One sensor changes the accuracy materially — a fluorometer at the intake itself, because it measures the plant's own water instead of the water 35 km up the coast. That is the plant's capital decision, and the free audit tells us whether it is needed.

What this is not

The ask is small and the first step is free

One plant, one dataset: 12–24 months of production, dosing, cleaning and turbidity records. We return a retrospective on what the last bloom actually cost that plant — free, under NDA, before anyone discusses a pilot.

See the live panel first →