A connected process laboratory
for bioenergy intelligence.

What executes locally

The Node service computes process states and stores runs on SSD. Three.js displays the resulting liquid volume, gas inventory, mixing and asset state. Playback and seek use the same simulation clock. Separate local APIs provide research-reactor records, industrial SCADA windows, retained forecasts and actual training histories.

Collection planning computes pickup mass, capacity and a geometric nearest-neighbour route. The city page allocates supply against demand and calculates a parameterised energy-value scenario. Image review supports human labels. Demo telemetry validates units and timestamps and retains records.

Four 48-hour experiments

These are simulated scenarios using explicit, uncalibrated parameters. They are not measured production or field-validation results.

Electricity includes consumption of the initial gas inventory. Comparison retains common initial states and duration. Each source run remains in the Experiment memory page.

Equations and assumptions

S is model organic inventory in kg VS-equivalent, F is incoming organic mass per hour, V is liquid m³, G is stored standard gas volume and T is °C. Mixing-off multiplies reaction rate by 0.65. Y = 0.35 Nm³ CH₄/kg converted inventory and x_CH4 = 0.60 are illustrative scenario parameters. The temperature curve and kinetics are not fitted to a plant.

Heat balance assumes water-like density/heat capacity and a 2 kW/K loss coefficient. Electricity assumes 9.94 kWh/Nm³ CH₄ and 35% electrical efficiency. The engine integrates 15-minute steps and tracks gas/organic ledger residuals; it does not claim complete elemental stoichiometry or full ADM1.

Scientific data and ML

DS-05: Mendeley black-liquor research-reactor dataset; 138 canonical rows and 86 derived methane observations. Reactor identities remain separate and absent measurements stay missing.

DS-06: University of Iowa Research Online / Muscatine WRRF; 500,400 source SCADA rows, 24,038 retained prediction identities and recorded LSTM epoch histories. Its target is total-biogas CFM. Local forecast metrics are recomputed from retained predictions.

The primary run configuration uses a 24-hour LSTM input; the separate common-target memory study selects 12 hours. This distinction corrects the initial lab summary. LSTM, XGBoost and persistence are shown in their actual experimental context, rather than relabelling retained forecasts as new training.

Current limits and field pathway

Physical plant connections: zero. CNN inference, live instrument calibration, plant-calibrated VFA/ALK sensing and actuator control remain to be integrated and validated. Gasification is a separate conceptual animation. The city schematic has illustrative locations and no pressure solver. External legacy portal links require internet.

Next field work: secure compatible instrument/lab data and image rights; implement calibrated read-only gateways; validate forecasts and uncertainty prospectively; assess site transfer; conduct operator acceptance before control integration.

Verification record

Sixteen automated tests cover process balances, replay, intervention effects, invalid inputs, API persistence, evidence counts, industrial/epoch access and eleven page workflows.

Unavailable: macOS computer-use permissions pending. DOM checks stub renderer; visual/WebGL/mobile acceptance not verified.

Source: GFIS_Next_Product_Lab_2026/03_Prototype_Implementations/product. Exact startup and replay commands are in 00_Project_Control/LOCALHOST_RUNBOOK.md.