AquaSentinel — RO fouling forecast and early warning

AquaSentinel recognizes a developing membrane-performance trajectory and forecasts when it may become a significant performance problem. It projects time to a documented intervention criterion so operators can inspect the affected train and decide what to do before that criterion is reached. Its open benchmark makes the forecast testable. Pick a fault and drag the severity to inspect one synthetic run.

Why it matters

Illustrative comparison of reactive intervention and an earlier forecast advisory. Operational value remains to be validated with plant data.

An earlier credible advisory creates a decision window. Plant evidence must determine whether that window improves an operator's decision.

Fault type

Severity — moderate

Advanced — plant & model parameters

Feed salinitymg/L Feed temperature°C Recovery% Clean permeability A₀LMH/bar Fault onsetday Sensor noise×

Cleaning-in-place (CIP) cycles on

Modeled intervention window — —

Detection delay after onset —

ROC-AUC (this run) —

Energy SEC, clean → criterion (kWh/m³) —

—

Modeled membrane fouling outlook —

Likely condition—

Expected criterion crossing—

Modeled intervention window—

Suggested operator response—

Risk colors use documented time bands, not an opaque AI score.

permeability (true) measured flow criterion 0.90 forecast advisory industry criterion

—

Detectors compared — this run

Detector	Role	Fires (day)	Lead vs criterion

A benchmark is a yardstick: every method sees the same labeled run and declared comparator.

Open data

Across the full benchmark (computed live in your browser · 3 seeds × 3 severities per fault)

Fault	Detection	Mean modeled window	ROC-AUC

The browser table is a small Gulf-preset demonstration. The locked Python protocol contains 300 held-out runs across three regimes: mean lead is about 1.3 days for gradual faults, with 0 observed process alarms in 120 synthetic controls. These are modeled results, not plant performance.

Evidence & reliability

Can you trust the method before it reaches a plant? The proof of concept combines quantitative results, physical plausibility checks, and a clear map of completed and pending validation.

Left: lead-time sensitivity to 12, 24 and 48 hour confirmation policies. Right: held-out lead distributions across three operating regimes. — The comparator uses DuPont-aligned 10%/10%/15% limits. Its confirmation duration is disclosed because it materially changes the modeled intervention window. The primary 24-hour policy gives about 1.3 days on gradual faults.

The modeled values sit inside the cited SWRO ranges. This is a plausibility check, not plant validation.

Validation ladder: literature plausibility, implementation checks, industry-aligned protocol and held-out synthetic testing are completed; design-tool and plant validation remain pending. — Honest status: what's verified now, and the real-data validation that comes next.

What earlier action could be worth — illustrative; set your plant's numbers

Plant capacity 500,000 m³/day Electricity price $0.080/kWh Grid CO₂ intensity 0.45 kg/kWh Fouling events / year 6

Extra energy avoided / event—

Cost avoided / event—

CO₂ avoided / event—

Annualized potential—