chemai

AI for process problems, built by a chemical engineer. Each project starts from a question a plant actually asks, lets the physics shape what the model sees, and says where its answer stops being reliable.

Eng. Mohammed Taleb · LinkedIn · Project pages

Distillation soft sensor

The lab checks the vapour pressure of the bottoms product about once a shift, and the result comes 1–3 hours after the sample. This estimates it from two readings the plant already has — one tray temperature and the column pressure — with a 95% range and a warning when the conditions are outside what it was tested on.

Try it live · Project page · Write-up · Validation record · Data exploration

Tested on a later operating campaign that ran at less than half the reboiler duty — never a random split:

Measure	Result
R²	0.980
Mean absolute error	0.645
Interval coverage (95% nominal)	94%
Error on points it flags as outside the tested conditions	7.8× the rest

What it found

• The physics changed the input, not the model. Writing temperature as 1/T, the form Antoine's equation uses, cut the error by about a third (MAE 0.953 → 0.645). A simple Ridge model then beat Random Forest and XGBoost, and two inputs beat all twenty-one (R² 0.980 against 0.884).
• Pressure needs temperature. On its own, pressure scores R² −0.50; next to temperature it is essential.
• The refusal is real. The envelope check was tested, not assumed: the points it flags have a 7.8× larger error.
• Its main limit: it under-predicts high values, so it is not an off-spec alarm.

AI for RO

One seawater RO plant, two goals: see its fouling coming, and run on the sun. Public data only — one real plant's published record, ten years of NASA sun and sea, public membrane datasheets — and our own models where the data stop.

The deck · Write-up · Notebooks

• No forecaster beat "no change" on the real record. Ten methods, from simple rules to deep learning, under one declared test: none forecast a train's ΔP 1–14 days ahead better than "no change".
• A model can still plan the cleanings. On simulated plants, reading only the routine columns any plant logs, it kept about the same number of cleanings and cut the days above the cleaning limit by 90 % — a simulation, not a real plant.
• A plant that follows the sun takes about 80 % of its power from it: twice the demand in size, with panels and a tank of product water, and no battery.
• Fouling costs little energy — about 1 % more on a typical day — and the sun still carries the solar plant.

Run it

pip install -e ".[dev]"
pytest tests -q            # 252 tests, no data needed
python -m chemai.api.run   # http://localhost:8000

curl -X POST localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{"temperature": 450, "pressure": 228}'

{"vapour_pressure": 38.8, "interval_lower": 35.84, "interval_upper": 41.75,
 "in_envelope": true, "distance": 1.26, "warning": null}

No endpoint returns a bare number: every answer carries its interval and the envelope check. The deployed API is at chemai-tr2c.onrender.com/docs — free hosting, so the first request after a quiet spell can take a minute.

Data is not committed. The soft sensor uses the public distillation-tower dataset from OpenMV.net.

What's in the repository

chemai/        the package: physics features, models, uncertainty, envelope check, API
projects/      one folder per project: training script and write-up
docs/          validation record and data exploration
tests/         252 tests that run without any dataset
soft-sensor/   the soft sensor's project page, served by GitHub Pages
ai-for-ro/     the AI for RO deck, served by GitHub Pages

How it is built

• Limits are published. Each tool states what it cannot do, next to the answer it gives.
• Negative results are kept. A model that lost to a simpler one is recorded with the number that shows it.
• Every error found became a test. test_coefficient_sign_is_physical fails if the model inverts against thermodynamics, test_split_is_not_random fails on leakage, and test_rejects_celsius_temperature catches the silent unit error.
• A model carries its provenance. The saved estimator records the library version that fitted it, and /health reports a mismatch.

These are deployed prototypes, not plant systems: no historian connection, no authentication, no monitoring, no automatic retraining, no field validation.

Implementation assisted by AI tools. The engineering framing, interpretation and validation are the author's.