Solutions · Companies
Can this recover value from my waste stream — and is it real?
MESSAI turns thousands of peer-reviewed studies into calibrated models so an industrial or commercial team can evaluate microbial electrochemical systems for wastewater treatment, resource recovery, and hydrogen — with honest uncertainty and a first-pass TEA — before spending capital on a pilot.
Grounded in the literature · honest uncertainty · no black boxes
What you get out of it
Outcome 01
See the recovery — and the honest economics
Enter your influent and get a predicted recovery balance (water, electricity, biogas, H₂, struvite, NH₃) with a mini-TEA over the plant life (15 years for MES trains, 25 for activated sludge) and LCA for your stream. Numbers are literature midpoints, not a contracted quote — you see exactly where each one comes from.
Outcome 02
De-risk before capex
Every prediction ships with conformal uncertainty bounds and an out-of-distribution flag. You see where the model is confident and where it is extrapolating — so a pilot decision rests on calibrated evidence, not a single point estimate.
Outcome 03
Standardize on evidence, not vendor claims
Recovery and performance ranges trace to thousands of peer-reviewed measurements with published calibration coverage. Compare architectures on a common, auditable basis instead of taking a supplier deck at face value.
Live demo · your stream
Drive it on a stream like yours
Start with a wastewater stream, get a ranked system suggestion, and see a real ML prediction with conformal bounds and an out-of-distribution flag — then a recovery balance, calibration receipts, and a mini-TEA over the plant life (15 years for MES trains, 25 for activated sludge).
Illustrative — not design numbers
Every value below is a literature-midpoint estimate with published uncertainty, shown so you can judge whether the recovery is real for your case. Site engineering adds cost on top; these are not a contracted quote or a stamped design. The calibration and scope-limit panels at the bottom show exactly where the models are — and aren't — trustworthy.
Section 1 · Influent characterisation
Define the wastewater stream
Six industrial archetypes or set parameters within their literature ranges. Each slider shows hard bounds, typical band, sub-ranges by facility type, and the active preset's target value.
Section 2 · System suggestion · lab catalog designs
Pick a coupled architecture
Ranked by influent fit using the 5-physics-family router. Every card is a digital twin from the lab catalog; click one to drive every downstream panel.
Loading 3D reactor scene…
Section 4 · Live prediction · /api/ml/predict
Live ML prediction
Typical error on the paper-disjoint benchmark (bes-benchmark-v1) is about 0.8 orders of magnitude (×6) for power density, so the bands are wide by design. Assumed operating point, not yet an input: 100 Ω external load, Nafion 117 membrane, Pt/C cathode at the anode’s area.
Section 5 · Recovery balance
What this system actually recovers
Daily recovered products from a first-principles balance on the influent and the live prediction. Band widths are not to scale; values marked “—” lack the inputs to compute.
- · COD removal from the archetype baseline (60 %).
- · Biogas at the theoretical (Buswell) CH₄ yield — an upper bound.
- · Electrode electricity: insufficient inputs (no power prediction).
Section 7 · Design recommendations
No change the corpus can vouch for yet
Each card is a fitted corpus effect with its n and 95 % CI. A change counts as recommended only when the CI’s lower bound clears a 5 % gain; the rest are shown as candidates whose direction the corpus hasn’t established. Stage one to see it as a ghost overlay in 3D.
live prediction unavailable — lifts shown as relative % only
Section 7b · Raise confidence
33 evidence gaps for MFC + Anaerobic Digestion
Each row is a design lever the corpus cannot yet quantify with high confidence, the study that would close it, and the literature search that looks for it.
- external resistance → power density128 papers · cross-paper · effect not distinguishable from zero · lowSuggested study: vary external resistance; hold temperature, pH, HRT, COD_in, anode material fixed; measure power density + coulombic efficiency; ≥ 3 conditions per study.openalex: external resistance "microbial fuel cell" power density
- pH → power density110 papers · cross-paper · effect not distinguishable from zero · lowSuggested study: vary anolyte pH; hold temperature, HRT, COD_in, R_ext, anode material fixed; measure power density + coulombic efficiency; ≥ 3 conditions per study.openalex: anolyte pH "microbial fuel cell" power density
- temperature → COD removal108 papers · cross-paper · effect not distinguishable from zero · mediumSuggested study: vary reactor temperature; hold pH, HRT, COD_in, R_ext, anode material fixed; measure COD removal + coulombic efficiency; ≥ 3 conditions per study.openalex: reactor temperature "microbial fuel cell" COD removal
- pH → COD removal105 papers · cross-paper · effect not distinguishable from zero · lowSuggested study: vary anolyte pH; hold temperature, HRT, COD_in, R_ext, anode material fixed; measure COD removal + coulombic efficiency; ≥ 3 conditions per study.openalex: anolyte pH "microbial fuel cell" COD removal
- influent COD → COD removal94 papers · cross-paper · effect not distinguishable from zero · lowSuggested study: vary influent COD concentration; hold temperature, pH, HRT, R_ext, anode material fixed; measure COD removal + coulombic efficiency; ≥ 3 conditions per study.openalex: influent COD concentration "microbial fuel cell" COD removal
- influent COD → coulombic efficiency62 papers · cross-paper · effect not distinguishable from zero · lowSuggested study: vary influent COD concentration; hold temperature, pH, HRT, R_ext, anode material fixed; measure coulombic efficiency + power density; ≥ 3 conditions per study.openalex: influent COD concentration "microbial fuel cell" coulombic efficiency
Section 7c · What to measure
Measurement plan for MFC + Anaerobic Digestion
Ranked from 263 MFC papers that report a recovery target: levers a study should vary, the context variables the corpus expects alongside them, and the outcomes to report. Coverage and correlation are descriptive; only the effect column is a fitted causal-grade estimate.
| parameter | role | reported in | fitted effect | gaps it closes | ρ vs targetpooled |
|---|---|---|---|---|---|
| anode material | lever | 0% · 0 | +0.017 → power density · low · n=5 | 64 · hold in 64 | — |
| hydraulic retention timeh | lever | 12% · 31 | −0.029 → COD removal · low · n=43 | 15 · hold in 113 | -0.25 (11) |
| pH | lever | 34% · 88 | −0.086 → ammonium removal · medium · n=10 | 14 · hold in 114 | +0.55 (15) |
| temperature°C | lever | 22% · 59 | +0.0087 → COD removal · medium · n=108 | 14 · hold in 114 | -0.25 (103) |
| COD Concentrationmg/L | lever | 33% · 87 | +4.6e-7 → COD removal · low · n=94 | 13 · hold in 115 | +0.43 (69) |
| External LoadΩ | lever | 29% · 75 | −0.0000037 → coulombic efficiency · medium · n=60 | 8 · hold in 120 | +0.78 (8) |
| Power DensitymW/cm² | covariate | 54% · 141 | — | — | +0.74 (30) |
| voltageV | covariate | 43% · 114 | — | — | +0.48 (13) |
| Open Circuit VoltageV | covariate | 35% · 91 | — | — | -0.45 (10) |
| Current DensitymA/cm² | covariate | 34% · 88 | — | — | +0.56 (26) |
| Reactor VolumemL | covariate | 32% · 83 | — | — | -0.20 (73) |
| Internal ResistanceΩ | covariate | 30% · 78 | — | — | -0.71 (12) |
| Cod Removal% | outcome | 83% · 219 | — | — | — |
| coulombic efficiency% | outcome | 44% · 116 | — | — | — |
| Ammonium-N Removed% | outcome | 8% · 21 | — | — | — |
| Phosphorus Removal% | outcome | 4% · 11 | — | — | — |
source: apps/web/public/data/computed/research/measure-priorities.json · 2026-10-07
Section 8 · Sensitivity ladder
Which knob moves the needle
Corpus effects from within-paper-effects.json, for this archetype. Within-paper fits partial out paper-level confounds; cross-paper fits don’t, and are labelled. Green raises, red lowers the target — but only where the 95 % CI excludes zero; grey bars are directions the corpus hasn’t established. β units differ by knob (per °C, per pH, per hour, per level), so compare signs and intervals, not bar lengths.
Mechanistic closure · direction-validated levers
Does the physics closure predict the direction (sign) of a within-design change better than a coin flip (50 %)? Only where the 95 % interval stays above 50 % — so far, substrate → power, narrowly. Direction agreement only, not magnitude.
Not shown: current↔substrate and current↔pH are anti-skillful (the closure predicts the wrong direction), so we exclude them rather than surface a misleading bar. The richer interactive version lives at /lab/design.
Section 9 · Calibration · honest caveats
What this model is — and isn't — calibrated for
⚠ Conformal coverage guarantee is provisional
The split-conformal q̂ were calibrated against holdout-priors-2026-05-21, but the served predictor is v2-sota-2026-05-15. The interval math (log-space + physical clamp) is correct, but the marginal-coverage guarantee holds only once the constants are regenerated against the served fit. Every /api/ml/predict response carries this same version_check.ok = false.
What we don't claim
- •MDC, MES, MNRC, MMRC, MBES are not in the held-out coverage check — analytical predictors only.
- •TEA numbers are straight-line literature midpoints, not a contracted quote; site engineering adds cost on top.
- •COD removal predictions are coarse for high-strength industrial streams with toxic shocks.
- •No predictions of micropollutant removal, antibiotic resistance, or pathogen fate.
- •Long-term biofouling, electrode degradation > 12 months are absent from the corpus.
Section 10 · Mini-TEA · MFC + Anaerobic Digestion
What it costs
From the archetype's Process-TEA baseline. Discount rate 0.1, lifetime 15 yr. Honest disclaimer: literature midpoints, not a contracted quote.
- · Electrode capex $1,200/m² (carbon-cloth, Logan 2008 inflated).
- · Performance inputs are stub literature midpoints — wire to a sweep export for real numbers.
- · Higher 10 % discount rate reflects MFC scale-up risk (Santoro 2017).
Every number traces to a source file · no black boxes · audit dashboard · full technical demo
Who it's for
Wastewater utilities & industrial dischargers
Turn a treatment cost centre into an energy- and nutrient-recovery opportunity — and quantify it before committing.
Resource-recovery & circular-economy teams
Screen struvite, ammonia, metals, and water reuse pathways against the evidence for a given feedstock.
Hydrogen & biogas developers
Assess MEC-H₂ and MFC/MEC + anaerobic-digestion routes for low-strength or recalcitrant effluent.
Process & TEA engineers
Get a defensible starting flowsheet and capex/opex band to build your own detailed model on top of.
Typical streams
The demo ships six commonly-cited industrial and municipal wastewater archetypes as starting points — each calibrated to published characterisations. Load one, or dial in your own.
Brewery effluent
High-strength, well-cited Logan anchor
Municipal wastewater
Largest TAM, low-grade energy + nutrient recovery
Dairy processing
Fats + proteins, balanced C:N
Landfill leachate
Toxic, recalcitrant — MEC-H₂ candidate
Blackwater (decentralized)
High nutrients, struvite + NH₃ recovery target
Pharmaceutical effluent
High-COD, micropollutant polishing
Get started
Bring us your stream. We'll show you what the evidence says it can recover.
Share an influent characterisation and target outcomes, and we'll walk you through a calibrated recovery + TEA assessment and where a pilot would de-risk fastest.
founders@messai.io