Physics + machine learning + forecasting + optimisation + safety

No black box: every layer, its inputs, outputs and method. Select a layer to inspect it.

Modeled

Architecture

LAYER 03

Learns what the lumped physics cannot represent; corrects its predictions.

Inputs: 16 features: what-if point, calibration point and conditions, θ̂, physics outputs.
Outputs: Residuals for permeate flow (relative), TDS (log-ratio), ΔP (bar), power (relative).
Method: scikit-learn HistGradientBoosting, 350 trees × 4 targets, trained on 40 000 synthetic samples; exported to JSON and evaluated in the browser; browser and scikit-learn predictions agree to within 1e-9.
Live: Waiting for telemetry…

ml/train.py

Loading model metrics…

training envelope extrapolation

Feed salinity36–4841.5 g/L

Temperature18–3628.4 °C

Feed pressure52–7363.5 bar

Feed flow / vessel6.5–149.6 m³/h

Membrane NPF87–100%97.0%

Distance to training distribution (Mahalanobis / d₉₉)0.00

0d₉₉1.4 · d₉₉ (withhold)2.5

Confidence

—

One train, what-if	Reference plant	Physics	ML only	Hybrid

Reference plant = the higher-fidelity simulator used as ground truth in this prototype (not a real plant); errors are relative to it.