Methods

Method

MFC (served with a stated COD) — polarization closure

mfc-mechanistic-closure-v1 · anodic_oxidation · MFC

Three MFC models are in use. The lab MFC tiles use mfc-butler-volmer-monod-v2. /api/ml/predict, the MCP server and the chat assistant use mfc-mechanistic-closure-v1 when a request states a COD and Enhanced_MFC_Experimental when it does not. For the same design they can differ. mfc-butler-volmer-monod-v2Enhanced_MFC_Experimental

What it predicts

Peak areal power, current and voltage and coulombic efficiency for a stated reactor, read from its polarization curve, plus the operating point under a stated external resistance. No COD removal.

Governing relations

Each relation is transcribed from the implementation. Where the code cites no source, the source line says so.

  1. Polarization curve
    V(j)=OCV−ηact(j)−jRint−ΔVpH,P=max⁡j j V(j)V(j) = OCV - \eta_{act}(j) - j R_{int} - \Delta V_{pH}, \qquad P = \max_j\ j\, V(j)

    Source: Python reference services/ml-engine/simulator/mes_anodic_oxidation.py, kept parity-tested; constants calibrated to the Logan-group MFC anchors.

    SymbolMeaningUnit
    VVcell voltageV
    jjcurrent density after the substrate limitA/m²
    OCVOCVopen-circuit voltage by cathode and cultureV
    RintR_{int}areal internal resistanceΩ·cm²
    ΔVpH\Delta V_{pH}pH overhead across a membraneV

    Worked example

    These inputs, passed to the predictor, return the output below. A test re-runs the example on every change, so this number is the one the code produces.

    inputvalueunit
    anodecarbon-clothmaterial
    cathodept-cmaterial
    temperatureC30°C
    ph7pH
    codMgL1,000mg/L
    anodeAreaCm27cm²

    Output · powerOutput = 301.92 mW/m², within ±0.1 %

    By hand · Mixed anaerobic culture, air cathode, no membrane: OCV 0.75 V, R_int 3100 Ω·cm², read off the curve maximum of a 20-point sweep. The closure is numerical; this pins its result, not a hand calculation.

    Reproduce · Pinned by libs/shared/ml/src/predictions/mfc-mechanistic.test.ts. Run it from the repository root:

    pnpm exec jest -c libs/shared/electrochemistry/jest.config.js mfc-mechanistic.test.ts -t "powerOutput"
  2. Internal resistance
    Rint=R0⋅m⋅(1+log⁡10A50)R_{int} = R_0 \cdot m \cdot \left(1 + \log_{10}\frac{A}{50}\right)

    Source: The scale-up rule was attributed to Logan 2008 §6.3; that section was not matched to the book in the 2026-10 citation audit, so it stands as unverified.

    SymbolMeaningUnit
    R0R_0base resistance for the stated reactorΩ·cm²
    mmmembrane multiplier1
    AAanode area above 50 cm²cm²
  3. Operating point
    V(j)=j A RextV(j) = j\, A\, R_{ext}

    Source: Load line on the polarization curve. Within-design, the direction of current against R_ext is right 71–78 % of the time; the location of the power peak is not (apps/lab design-sensitivity.ts).

    SymbolMeaningUnit
    RextR_{ext}external resistanceΩ

Assumptions and constants

  • Mixed anaerobic culture and an air cathode unless stated; the receipt names them as defaults.
  • Reactor constants are calibrated on Logan-group anchors; scores on reactors they were not tuned on are in docs/evals/published-anchors.json.
  • Materials the closure does not tell apart scale R_int by the material table ratio to carbon cloth with Pt, itself hand-calibrated.

Corrections applied

  • Monod, pH and Arrhenius factors on the exchange current.

Not modelled

  • COD removal.
  • Electrode spacing.

Literature anchors in the source

  • Liu & Logan 2004, Environ. Sci. Technol. 38:4040
  • Logan et al. 2007, Environ. Sci. Technol. 41:3341
  • Picioreanu et al. 2007, Water Res 41:2921

Implementation

libs/shared/electrochemistry/src/mechanistic/anode-model.ts

Interval the predictor ships

heuristic_fixed_fraction — A fixed ±25 % placeholder band. It is NOT a calibrated interval; when served through /api/ml/predict it is replaced by the benchmark band below, or withheld. How served intervals are built

Calibration artifacts used when served

  • Benchmark out-of-fold residual quantiles: bes-benchmark-v1 · gkf5 · generated 2026-09-09
  • Split-conformal calibration: split-conformal · generated 2026-05-21
  • Hierarchical-priors v2 trust snapshot: populated · generated 2026-05-16

Benchmark cells for this predictor’s systems

Each row is a stratum with n ≥ 30 residuals at nominal level 90%. Coverage is the cross-validated share of held-out points inside the band.

targetstratumnpaperscoverageCV
cod_removalMFC60217288.5%
cod_removalMFC|bioremediation46856.5%
cod_removalMFC|power_generation1344188.8%
cod_removalMFC|unknown53885.7%
cod_removalMFC|wastewater_treatment36011387.8%
coulombic_efficiencyMFC3019789.7%
coulombic_efficiencyMFC|power_generation1073186.9%
coulombic_efficiencyMFC|unknown301073.7%
coulombic_efficiencyMFC|wastewater_treatment1354877.0%
current_density_arealMFC49613489.3%
current_density_arealMFC|power_generation2085488.0%
current_density_arealMFC|unknown491483.7%
current_density_arealMFC|wastewater_treatment2015785.6%
power_density_arealMFC87124588.9%
power_density_arealMFC|bioremediation49969.4%
power_density_arealMFC|power_generation34311489.5%
power_density_arealMFC|unknown1062485.8%
power_density_arealMFC|wastewater_treatment3509185.4%

Prior trust by served parameter

The state the lab badge shows for each served parameter and system, from the v2 priors fit.

parametersystemstatepapersprior
power_density_arealMFCcurated275weakly_informative
current_density_arealMFCcurated150weakly_informative
coulombic_efficiencyMFCcurated96literature
cod_removalMFCcalibrated183weakly_informative

Scientific-rigor gates

pnpm verify:science runs these rules over the committed priors and validation artifacts in pre-commit. Known violations are baselined per rule and slug; a new one blocks the commit. Baseline from hierarchical-priors-v2.json@2026-10-28T19:44:43.057561+00:00.

rulebaselined
gate-honestyNo parameter may pass a rounded-threshold gate read while converged=false0
support-containmentBounded quantities must have ci95 (and pi95 when present) inside physical support, width <= range34
magnitude-sanityNo ci95 bound may exceed the plausibility ceiling (1.0e+6)0
slug-uniquenessNo two slugs may share a spelling-independent identity; scale must agree1
secondary-source-purityNo fitted prior may carry a REVIEW (secondary-source) stratum0
artifact-pairingEach fit must have its own coverage evidence; the served fit must have some0
coverage-sharpnessAny coverage figure must ship its interval width alongside0
interval-calibrationHeld-out interval coverage per parameter may not newly degrade76
effects-gap-completenesseffects.json is schema 2.0; every non-high effect has a gap per archetype; no gap is "high"0