AquaTwin

AquaTwin is a digital twin for seawater reverse-osmosis (SWRO) desalination. It combines a reduced-order physical model of the plant with a machine-learning residual that corrects what the physics misses, calibrates itself continuously against telemetry, forecasts membrane degradation, and recommends operating strategies that a deterministic safety layer, AquaGuard, must approve before they are shown to an operator.

Built by Team Kanban for the Khalifa University–UNESCO Global Water Hackathon 2026, theme Smart, Digital, and AI-Enabled Water Systems.

This is a research prototype. The plant it monitors is simulated: a higher-fidelity reference plant model stands in for a real SWRO facility so that ground truth is known and every claim can be tested. It is not connected to any real plant and is not intended for operational use.

Simulated

Plant behaviour produced by the reference plant simulator — the stand-in for a real plant in this prototype. All live telemetry is simulated.

Modeled

Outputs of AquaTwin's hybrid model (physics + ML residual), e.g. predictions, expectations and forecasts.

Estimated

Derived quantities with material outside uncertainty, e.g. carbon from an average grid intensity.

Assumed

Design choices and parameters without a specific source (flagged in the code and model documentation).

External ref.

Values taken from published sources: membrane datasheets, seawater property correlations, WHO guidance, grid data.

Measured

Not used. AquaTwin has not been connected to a real plant; no value in this prototype is a plant measurement.

No organisation named in this prototype or its documentation — including utilities, plant operators, equipment manufacturers, universities or UN agencies — has reviewed, endorsed or partnered with AquaTwin. Names appear only to cite published sources or to describe the hackathon.

Performance figures are simulation results or model estimates and are labelled as such. They have not been validated on an industrial plant. AquaTwin does not use a language model for any prediction, control or safety decision.

The prototype in 60 seconds, recorded from this application running on the simulated plant; every result shown is simulated. By Team Kanban. Voice: Kokoro-82M. Music: “Happy Beats / Business Moves Vol. 12” by ende.app. Sound effects: Kenney (CC0).

• Seawater osmotic pressure and density
  Sharqawy, Lienhard & Zubair (2010), as implemented in WaterTAP
• Solution–diffusion RO transport, 0D/1D structure
  WaterTAP reverse osmosis models
• Membrane element data, design limits, normalisation
  DuPont FilmTec™ SW30 datasheets and technical manual
• Standardised (normalised) RO performance
  ASTM D4516-19a
• Isobaric energy recovery
  Energy Recovery, Inc. PX Q400 white paper (2025)
• Physics-guided machine learning
  Willard et al. (2023), ACM Computing Surveys
• Conformal prediction intervals
  Angelopoulos & Bates (2023)
• Out-of-distribution detection (Mahalanobis)
  Lee et al. (2018), NeurIPS
• Gradient-boosted trees
  scikit-learn HistGradientBoostingRegressor (Pedregosa et al., 2011)
• Drinking-water quality context
  WHO, Safe drinking-water from desalination (2011)
• Grid carbon intensity (UAE, lifecycle)
  Ember via Our World in Data
• Harmful algal blooms and desalination
  IOC-UNESCO Manuals and Guides No. 78 (2017)

The full reference list, with notes on what was and was not verified, is in docs/REFERENCES.md.

• Next.jsMIT
• ReactMIT
• three.jsMIT
• Tailwind CSSMIT
• ZustandMIT
• d3-scale · d3-shape · d3-arrayISC
• MotionMIT
• Geist and Geist MonoSIL OFL 1.1
• scikit-learnBSD-3-Clause
• NumPy · pandasBSD-3-Clause
• TypeScript · tsxApache-2.0 · MIT
• Playwright (QA)Apache-2.0

Concept, engineering, modelling and design: Team Kanban.

Prototype version 0.1 · synthetic data seed 20261101 · all experiments reproducible from the scripts in the repository.