MESSAI.iO

Messai × Helmholtz-UFZ

Use case #1 · nitrogen cycling

Nitrogen cycling in bioelectrochemical systems, from corpus to designed experiment.

One end-to-end example that every later use case reuses, held to the standard of a co-authored paper. This edition shows what is measured today, what each stage will add, and what we need from UFZ.

Prepared for Dr. Benjamin KorthCorpus measured 2026-10-05 15:10 UTC← Collaboration briefingSource papers →

Cohort papers

490

nitrogen × BES

Full-text PDFs

164

33% of the cohort

Modelable facts

1,025

distinct, before QA

Complete records

—

pending step 5

01 · Where it stands

Integrity first, then the cohort

The pipeline runs in order. Nothing is extracted or fitted until the fixes that keep data from being silently corrupted are in.

  1. 0

    One spec file per use case

    Every stage reads the use case from one JSON spec instead of hard-coded MFC/MEC lists. Today’s pipeline became instance #0 with byte-identical outputs; nitrogen is instance #1, and Kolbe or MES-reactor work becomes a new file, not new code.

    Merged 2026-10-02

    Done
  2. ·

    Integrity fixes before any new extraction

    Re-syncs archive superseded rows instead of deleting and re-inserting them; values cited from other papers are gated out; nitrogen units carry their species basis (1 mg NO₃⁻ is 0.23 mg NO₃⁻-N); N₂O gets its own parameter; nitrogen parameters get physical bounds; wrong aliases are corrected (nitrate is an electron acceptor, not a donor).

    Archive-not-delete helper built; wiring next

    In progress
  3. 1

    Freeze the cohort

    The query below runs once and writes manifest v1: paper ids, DOIs, query text, database, date and sha256. Later additions, such as UFZ-supplied PDFs, become v1.1 with a logged diff.

    Next
  4. 2

    Acquire full text

    The free open-access chain is exhausted for most of the cohort’s missing PDFs. The rest needs institutional or text-and-data-mining access.

    Needs UFZ
  5. 3

    Parse and extract, gated

    Full-text extraction with the cited-value gate and a review gate. A 5-paper smoke run comes before any scale-up.

    Later
  6. 4

    Sync and harmonise

    Canonical parameters, SI units with species basis, physical bounds and the verifier; system types reclassified with the canonical taxonomy.

    Later
  7. 5

    Experiment records and folds

    Condition sets graded as complete records. Paper-disjoint folds, with fold 0 locked before any fit.

    Later
  8. 6

    Insight

    Effects for each hypothesis, evidence gaps, and nitrogen priors refit per system class.

    Later
  9. 7

    Recommendation

    Operating windows per knob and a ranked experiment design with a predicted value and interval for every run.

    H5 worked example: run sheet pre-registered on this page

    In progress
  10. 8

    Design

    Denitrification physics with the N₂O emission factor as the headline output, a corrected NH₃-recovery model, a sizing dossier, P&ID and 3D.

    Later

02 · What the corpus holds today

490 papers in, no complete records out — yet

The cohort query below, run against the Messai corpus. Counts are distinct facts: duplicate copies from earlier syncs are collapsed, so nothing is counted twice.

Browse all 490 source papers — corpus pages, PDFs and DOIs →

How a paper gets into the cohort

One rule, applied by a query. Nobody picks papers by hand. A paper is in when its title or abstract names a nitrogen topic and a bioelectrochemical system.

  1. 1

    Start with every live paper

    19,871 papers in the Messai corpus. Archived papers, and duplicates merged into another record, are left out.

  2. 2

    Keep those that name a nitrogen topic

    The title or abstract must contain one of these as a whole word:

    • nitrate
    • nitrite
    • ammonium…
    • denitrif…
    • nitrification…
    • anammox…
    • ammonia removal
    • ammonia recovery
    • nitrogen removal
    • nitrogen recovery
  3. 3

    …and a bioelectrochemical system

    It must also contain one of these:

    • microbial fuel cell
    • microbial electrolysis
    • microbial electrosynthesis
    • microbial electrochem…
    • bioelectro…
    • electroactive…
    • exoelectrogen…
    • electrotroph…
    • biocathode…
    • bioanode…
    • MFC
    • MEC
    • BES
    • MET

Terms are matched in English, as written in the paper. “…” stands for any word ending.

490 papers pass both tests. They are the cohort.

What the rule does not check

  • It reads the title and abstract only, never the full text.
  • A match means the topic is mentioned, not that it is the paper’s subject. A paper that names nitrate only as a medium ingredient can still qualify. Every paper is listed, so each can be checked.
  • The six sub-process labels below are tags on papers already in the cohort. They never add or remove a paper.

Corrected 5 October 2026

The first version matched terms as fragments of words and counted “groundwater” alone as a nitrogen topic. That let in 59 papers that do not study nitrogen: 12 whose only match was a reagent name (for example “cetyltrimethylammonium bromide” in a hemoglobin paper) and 47 about groundwater cleanup of other pollutants. Both are fixed, and the cohort went from 549 to 490 papers.

Papers

  1. Cohort papers490

    nitrogen topic AND bioelectrochemical, title + abstract

  2. With a DOI48499%

    resolvable for acquisition

  3. With a full-text PDF16433%

    the rest is mostly paywalled

  4. With any extracted values27155%

    includes abstract-level and legacy extraction

  5. With current-extractor values398%

    v2 full-text extractor

Extracted facts

  1. Extracted facts4,376

    distinct values, duplicate copies collapsed

  2. Mapped to a canonical parameter2,59159%

    has a canonical slug

  3. Modelable1,02523%

    SI-normalised and flagged usable

  4. Verified nitrogen facts00%

    no nitrogen parameter has physical bounds yet

  5. Complete experiment recordsPending

    all required inputs + an output in one condition set

By sub-process

A paper can count in more than one row. Click any value for what it counts.

Sub-processPapersWith PDFWith valuesModelable facts
papers →
papers →
papers →
papers →
papers →
papers →

Nitrogen parameters across the whole corpus

Not only the cohort. An amber 0 means none of that parameter’s facts normalise to SI units yet (no unit kind, or unparsed spellings), so no stage downstream can use it. Click a parameter or value for its definition, why it reads as it does, and its model status.

ParameterFactsModelablePapersVerified

SYSTEM TYPE · MFC 238 · MEC 38 · REVIEW 13 · MES 10 · MSC 10 · MDC 9 · unclassified 147

Where the data is soft — stated up front

  • 6,351 extracted rows hold 4,376 distinct facts: 1,975 (31%) are duplicate copies from earlier syncs. An audited de-duplication runs before the cohort is frozen.
  • 0 nitrogen facts are verified, because no nitrogen parameter has physical bounds yet. Bounds come with the integrity fixes.
  • 960 facts across 284 papers name a nitrogen quantity but map to no canonical parameter, and N₂O has no parameter at all.
  • 147 of 490 cohort papers have no system type, and most of the rest are labelled MFC. They are reclassified with the canonical taxonomy before any per-class fit.
  • 76 of the 1,025 modelable facts come from a legacy bulk import that is not yet grounded against full text. They are re-checked before use.

03 · Three base records, one cycle

What counts as one usable experiment

A record is one condition set with every required input and at least one output. The whole nitrogen cycle is a coupling of these records within the same reactor, not a fourth model.

Cathodic denitrification

239 papers

Required

  • Influent nitrate (as N)
  • Cathode potential vs SHE, or applied current density
  • HRT

Covariates

  • Temperature
  • Catholyte pH
  • COD : N ratio

Outputs

  • Nitrate removal (rate or fraction)
  • N₂O emission factor
  • Nitrite accumulation
  • Faradaic efficiency to N

Nitrification

97 papers

Required

  • NH₄⁺ (as N)
  • Dissolved O₂, or anode potential vs SHE
  • HRT

Covariates

  • Temperature
  • pH

Outputs

  • NH₄⁺ removal (rate or fraction)
  • Nitrite / nitrate formed
  • N₂O production rate

NH₃ recovery

37 papers

Required

  • NH₄⁺ (as N)
  • Applied current density
  • HRT or run duration

Covariates

  • Catholyte pH
  • Membrane type

Outputs

  • NH₃ recovery
  • NH₄⁺ flux across the membrane
  • Energy per kg N

04 · See it

Models, P&IDs and the protocol behind each record

The 3D models that exist today for each base record, with the process sheet where one is drawn, and the nitrogen steps now in the shared protocol graph. Drag a model to turn it.

BES reference geometry

Microbial electro-remediation cell

The bioanode oxidises organics; electrons cross the external circuit to a remediation biocathode that reduces nitrate to N₂, shown rising as bubbles. This is the cell the design step will size against the drinking-water limit.

Microbial electro-remediation cell

The nitrogen protocol

The steps added to the /protocols graph for this use case, on the generic backbone they hang off. Left to right is execution order; pick MERC or MNRC top-left to see one system’s path.

Loading the protocol graph…

Click a step for its measurements, equipment, precautions and references. Drag to pan.Open the full protocol graph →

Not modelled in 3D yet: a nitrifying bioanode, and the denitrifying column the design dossier will size. Both arrive with step 8.

05 · Hypotheses to sign off

Six confirmatory questions, fixed before the hold-out opens

Everything else is exploratory and reported with Benjamini–Hochberg false-discovery control. The direction and minimum effect worth acting on are UFZ’s call. The evidence columns are what the corpus holds for each one today. Click a row for the mechanism, the data behind it and what is pending.

#RecordDriver → outcomePapers with bothBoth variedEvidence todayStatus
73
73
00
32
148
00

“Papers with both” counts primary papers that report the driver and the outcome — reviews are excluded because their values are mostly cited. “Both varied” counts those reporting at least two values of each, the minimum for a within-paper direction. Untestable = none; anecdotal = under 5; thin = under 15. Click a count to open those papers. Only H5 has enough to work through, below.

06 · H5, end to end

From the corpus to a pre-registered experiment

The one hypothesis the corpus can begin to inform, taken through every stage: what the papers say, what physics allows, what to run next, and what we predict before running it.

Step 1

What the corpus can say about it

14 cohort papers report both a current and an NH₄⁺ outcome, but only 8 vary both and 9 pair them in the same condition set (measured 2026-10-05 15:10 UTC). Most report removal as a percentage rather than a flux. That is enough to see a direction, not to estimate an effect. The clearest within-paper series:

Combined anaerobic–aerobic bioelectrochemical system for ammonium, sulfide and carbon removal from ethanol stillage · J. Chem. Technol. Metall. 58(2), 2023 · corpus page · PDF ↗ · doi:10.59957/jctm.v58i2.53

ModeCurrent density, A m⁻²NH₄⁺ anode → cathode, mg L⁻¹Removal (Table 3)Text says
No BES (control)—166 → 77.0653.6 %53.58 %
MFC0.01–0.02155 → 48.568.7 %67–69 %
MEC 0.6 V0.4–1.1139.1 → 45.567.3 %67–69 %
MEC 0.8 V0.8–1.5125 → 30.575.6 %no number in quote
MEC 1.0 V1.5–2.9108.9 → 9.990.9 %90.91 %

Removal climbs from 54 % with no circuit to 67–69 % for the MFC and MEC at 0.6 V, 76 % at 0.8 V and 91 % at 1.0 V — one paper, one direction. Every check the page can run on it:

  • PassStated removal vs Table 3. Every removal the text states matches (anode − cathode) / anode from Table 3 within 1 percentage point.
  • FlagEvery value appears in its quote. The extractor recorded 75 % for 0.8 V (MEC), but its quote only says “followed by MEC0.8V” — no number. Table 3 gives 75.6 %.
  • FailPower ÷ current gives a possible voltage. 50.1 W m⁻² at 133.6 mA m⁻² implies 375 V across one cell — impossible (a single cell stays near or below ~1 V). Most likely the power is 50.1 mW m⁻², which gives 0.38 V.
  • Can’t tellRemoval attributable to a mechanism. Removal is measured across an aerobic cathode, where nitrification, NH₄⁺ migration and NH₃ stripping all remove ammonium. Without nitrite/nitrate in the cathode effluent, the electrode area and the flow, the share due to migration cannot be separated.
  • BoundFaraday ceiling on NH₄⁺ migration. At 1.5–2.9 A m⁻² (MEC 1.0 V), migration can move at most 18.8–36.4 g N m⁻² d⁻¹, even if NH₄⁺ carried every charge.

Step 2

What physics allows

JN = tNH₄ · i / F · MN · 86 400 ≤ 12.54 g N m⁻² d⁻¹ per A m⁻²
tNH₄ = DNH₄cNH₄ / Σ zj²Djcj (cations only)

Current sets how much charge crosses the membrane; the NH₄⁺ share of the cations sets how much of it is ammonium. That splits the design question in two: energy per kg N does not depend on current density in this model. It is the cell voltage times 24 / (t × 12.54), so about 5.7 kWh kg⁻¹ N per volt at a 30 % NH₄⁺ share and 2.6 at 70 %. In a real reactor, voltage climbs with current, so the energy cost of pushing harder shows up through the cell voltage. The run sheet measures exactly that.

  • Ideal cation-exchange membrane: anions carry no current.
  • Solution diffusivities (CRC, 25 °C) stand in for membrane-phase mobilities.
  • Migration only — back-diffusion, NH₃ volatilisation and water transport are ignored, so every prediction is a ceiling. Nothing is fitted to corpus data.

Step 3

The experiment, pre-registered

2² factorial + 3 centre points, current density × NH₄⁺ share of cation charge, run in the randomised order below. Current spans 2.5–10 A m⁻², from about the paper’s highest daily current upwards; the NH₄⁺ share spans 0.3–0.7 so its effect is measurable. Competing cations: K⁺ : Na⁺ = 1 : 1 by charge. Intervals propagate current density ±5 %, NH₄⁺ share ±10 %, diffusivities ±5 % (1σ, relative) through 4,000 draws.

Orderi, A m⁻²NH₄⁺ sharet NH₄⁺Predicted J ceiling, g N m⁻² d⁻¹ (90 %)kWh kg⁻¹ N per V
110.00.30.3442.5 (34.7–50.7)5.65
22.50.30.3410.6 (8.7–12.6)5.67
3centre5.00.50.5433.9 (28.3–40.2)3.52
42.50.70.7423.0 (19.3–27.1)2.60
5centre5.00.50.5434.0 (28.2–40.2)3.53
610.00.70.7391.9 (76.8–107.8)2.61
7centre5.00.50.5434.1 (28.0–40.3)3.52

Measure in every run

  • NH₄⁺, NO₂⁻, NO₃⁻ in both chambers and the acid trap (ion chromatography)
  • Na⁺, K⁺, Ca²⁺, Mg²⁺ in the feed — the actual NH₄⁺ share
  • Current and cell voltage, logged; membrane area; flow
  • A closed nitrogen balance (in = out across every species)

Fixed before the first run

  • Flux rises with current density and with NH₄⁺ share.
  • No observed flux exceeds its 90 % ceiling; one that does means the area or the measurement is wrong.
  • Observed ÷ predicted is reported per run as the apparent migration efficiency — the number the corpus cannot give today.

Pre-registration · run sheet SHA-256 1ac828f157891efe… · seed 20261003 · any change to the design changes this fingerprint.

Step 4

What would change our mind

Fluxes far below the ceiling at a high NH₄⁺ share would mean back-diffusion or NH₃ loss dominates, and the nitrogen balance would show where the ammonium went. Flux that does not rise with current would point at a competing cation the feed analysis missed. Either way the result becomes the first calibrated migration efficiency for this model — and a recorded paired series the corpus is missing. See the recommendations walk-through for how the same paper reads to its own authors.

07 · How every claim gets tested

Four stages, each on data the model never saw

The best-covered nitrogen parameter appears in 72 papers, so the locked fold holds only a handful of papers per outcome. Cross-validation carries the statistical power; UFZ data and prospective runs carry the claim.

  1. 1Development

    Not run

    Paper-disjoint 5-fold cross-validation on folds 1–4. Exploration here is unrestricted.

    Pass: Skill means beating the pooled-prior baseline, with a paper-bootstrap 95% CI on the difference that excludes 0. Otherwise the page says “no skill”.

  2. 2Locked hold-out

    Not drawn

    Fold 0, assigned by hashing the paper id with a salt fixed in the manifest. Nothing fits, tunes or selects on it; it is opened once, after hypotheses and models are frozen.

    Pass: Exchangeable (adversarial AUC ≤ 0.6). 90% intervals reach 0.85–0.95 empirical coverage.

  3. 3External

    Awaiting data

    UFZ datasets through a CSV intake template. Any UFZ-authored paper already in the cohort moves here, so nothing leaks.

    Pass: Same skill and coverage bars, on data the corpus never saw.

  4. 4Prospective

    Awaiting design

    UFZ runs the recommended experiments. Settings, predictions and 90% intervals are hashed and committed before each run, then scored.

    Pass: Observed value inside the interval, and the effect direction correct.

08 · What this page will show

Results land here as their stage completes

“No skill”, “not fitted” and “no recommendation” are first-class results, shown as such.

Insights

Pending · Step 6

Effect sizes for H1–H6, ranked evidence gaps, and nitrogen priors per system class. No nitrogen prior has converged yet, so none is shown rather than a fabricated interval.

Operating windows

Pending · Step 7

A window per knob (cathode potential, HRT, current density, C : N, pH), constrained by the maximum N₂O emission factor, nitrite accumulation and energy per kg N. Below the evidence threshold the answer is “no recommendation”.

Experiment design for H1–H4 and H6

Pending · Step 7

The same treatment as H5 above — run sheet, predictions with 90 % intervals, measurement plan, fingerprint — for the other hypotheses once the denitrification model exists. H3 and H6 first need papers that report their driver and outcome together.

Reactor design dossier

Pending · Step 8

A denitrifying biocathode sized from flow and influent nitrate to the EU drinking-water limit of 11.3 mg NO₃⁻-N/L (Directive (EU) 2020/2184), reported as a range from the interval, with a P&ID and 3D model. Until UFZ picks a case it uses an illustrative one: 10 m³/d at 25 mg NO₃⁻-N/L.

09 · What we need from UFZ

  1. Sign off the records and hypotheses. Check the three record definitions, confirm H1–H6, and set the constraints: maximum N₂O emission factor, nitrite accumulation and energy per kg N.
  2. Full-text access. 326 cohort papers have no PDF, and most of those have no open-access copy. Institutional or text-and-data-mining access (UrhG §60d / EU DSM Art. 3), confirmed with UFZ’s library first.
  3. Datasets for the external test. Which published UFZ datasets can be shared, in our CSV intake template.
  4. Design target. Groundwater column or H-cell, with the target flow and influent nitrate.
  5. Authorship and data release. Co-authorship and release terms for the paper.

Corpus numbers: scripts/collab/nitrogen-readiness.sql (sha256 c76b9991ab52), run read-only against the staging corpus on 2026-10-05 15:10 UTC. Cohort = live papers whose title or abstract names a nitrogen process and a bioelectrochemical system; facts are distinct (paper, parameter, value, unit, snippet, conditions). Corpus-wide: 19,871 live papers, 126,160 distinct facts, 21,519 modelable. Design: nitrogen-cycling golden standard, 2026-10-01. Prepared for Dr. Benjamin Korth, Helmholtz Centre for Environmental Research — UFZ.