Research · Build · DiscoverComputational Science & Engineering

Applied Computing

AirWater AI

Evidence-aware site screening for atmospheric water harvesting.

Is water-from-air worth testing at a given site, and if so, which material and operating schedule should be tested first?

AirWater AI screens whether harvesting water from air is worth testing at a specific site, and with which material and daily schedule.

Metal-organic frameworks (MOFs) can capture water vapor from air, but the material with the highest laboratory capacity is not always the one that produces the most usable water in a given climate. AirWater AI matches measured material behavior to a site's hourly climate and available heat.

Data & Validation

Leave-one-MOF-out test · each material is predicted from the other eleven

Measured data points
1,226
28 experimental water isotherms for 12 MOFs, from 13 published papers
Lower error than baseline
14%
Held-out error of 0.119 vs 0.138 kg/kg (mean absolute error)
Held-out R²
0.24
Up from −0.10 for the baseline; transfer to an unseen material remains limited

Three further exploratory materials have no measured isotherm and are excluded from training. The model improved on a baseline, but transfer to an unseen material remains limited, so predictions are screening signals.

How It Works

The same inputs can end in a recommendation or a refusal.

  1. Describe the site

    Location, month, daily water target, material mass, and available heat.

  2. Build the climate

    An hourly profile of temperature, humidity, and solar availability from NASA POWER data.

  3. Predict uptake

    Measured isotherms give each material's water uptake; a model fills only small gaps.

  4. Rank and schedule

    Candidates are ranked by yield, climate fit, heat feasibility, and evidence, with a 24-hour capture and release plan.

  5. Decide

    Seven checks end in a recommendation, a caution, or a do-not-deploy verdict.

Same site, two seasons. Phoenix, Arizona, with a 3 L/day target, 8 kg of material, and solar heat up to 176 °F. Only the month changes.

JulyAbove idealized screening threshold

AirWater AI top candidate for Phoenix in July: number 1, Al-Fumarate, 3.47 to 4.50 L/day-equivalent. Working capacity 0.277 kg/kg, regeneration target 167 °F, evidence 78%. Tags: Above idealized screening threshold, Best overall fit, Real isotherm.
Al-Fumarate ranks first, at an estimated 3.47–4.50 L/day-equivalent before device losses. The whole range is above the 3 L/day target.

JanuaryDo not deploy

AirWater AI top candidate for Phoenix in January: number 1, Al-Fumarate, 2.12 to 2.74 L/day-equivalent. Working capacity 0.169 kg/kg, regeneration target 167 °F, evidence 78%. Tags: Below idealized screening threshold, Best overall fit, Real isotherm.
Al-Fumarate still ranks first, but the estimate falls to 2.12–2.74 L/day-equivalent, below the target, and three of the seven decision checks fail.
How the candidates compareThe July run compares 15 materials and ranks the top three, each with the reason it is not first. In the chart every bubble is one material, placed by working capacity (kg/kg) against the temperature needed to regenerate it; bigger bubbles mean more simulated daily yield, and darker ones stronger evidence. CuBTC and MOF-303 hold more water than Al-Fumarate, but need more heat than the 176 °F this site can supply, so they sit right of the dashed line. Swipe sideways to see the whole screen.
AirWater AI decision view for Phoenix in July, headed 'Top candidate comparison'. Three ranked cards: number 1, Al-Fumarate, 3.47 to 4.50 L/day-equivalent, working capacity 0.277 kg/kg, regeneration target 167 °F, evidence 78%, tagged above idealized screening threshold and best overall fit. Number 2, SIFSIX-Zn, 3.74 to 4.14 L/day-equivalent, working capacity 0.091 kg/kg, regeneration target 149 °F, evidence 67%; why not first: lower climate fit and lower evidence. Number 3, MIL-160, 2.94 to 4.19 L/day-equivalent, working capacity 0.248 kg/kg, regeneration target 167 °F, evidence 80%; why not first: lower climate fit and higher uncertainty. Below, a bubble chart of 15 materials plots working capacity against regeneration temperature, with a dashed line at the user heat limit of 176 °F. Al-Fumarate and MIL-160 sit left of the line at 167 °F; MOF-303 at 185 °F and CuBTC at 221 °F have the highest working capacity, about 0.3 kg/kg, but sit right of it.
Why the month mattersAl-Fumarate's measured water uptake climbs steeply between about 20% and 40% relative humidity. In July the night-time capture window sits at 32%, giving a working capacity of 0.277 kg of water per kg of material. In January it sits at 26%, lower on the same curve, and the working capacity drops to 0.169 kg/kg: about 40% less water from every cycle. The screen below is the July run, with uptake in kg/kg plotted against relative humidity. Swipe sideways to see the whole screen.
AirWater AI evidence panel for Al-Fumarate in the July run, headed 'Evidence: is this interpolation or extrapolation?'. A chart of water uptake against relative humidity shows 59 measured points from the NIST ISODB database and an S-shaped fitted curve: about 0.05 kg/kg at 15% relative humidity, about 0.4 kg/kg at 45%, then level near 0.42. A shaded band marks the humidity the site sees across the day, roughly 18% to 35%. Adsorption point used: 32% RH, 0.284 kg/kg. Regeneration point used: 6% RH, 0.007 kg/kg. Both are within the measured range of 0 to 89% RH, and the model is marked as interpolating. Source: DOI 10.1021/acs.iecr.5b03509.

Real screens from the public application, cropped and otherwise unedited. All outputs are simulated research estimates, not certified device performance.

AirWater AI is a research-screening tool: it does not guarantee daily yield, potable water, lifecycle cost, or field safety.

What I Did

I designed and built AirWater AI as a functional proof of concept, from the data pipeline to the web application, and released the source code, data provenance, and model results.

  • Built the offline pipeline from NIST isotherm data to a traceable dataset
  • Fitted uptake curves and scored the evidence for each material
  • Trained and validated the water-uptake model
  • Wrote the site-specific ranking and scheduling logic
  • Added a refusal gate with seven decision checks
  • Built and deployed the web application

Web Application

Public screening tool

Open the app

Open-Source Software

airwater_app repository

View on GitHub