Which parameters drive Kolbe performance, and by how much

The question the predictor answers

query    feed = { acetic 0.30, butyric 0.20, caproic 0.50 } mol/L   (example inputs)
         anode = Pt · j = 50 mA/cm² · one-chamber · 20 °C · pH = ?
answer   CE_fuel   { value: pending, unit: "%",     ci_low: pending, ci_high: pending }
         energy    { value: pending, unit: "kWh/L", ci_low: pending, ci_high: pending }
drivers  total concentration · composition · current density · electrode   (ranked: pending)
evidence Pohl 2026 runs + Kolbe literature (extracted, review and harmonisation pending)
missing  pH / degree of neutralisation  →  named as missing, no default filled in, answer flagged

How this works

The path from a published paper to a number on this page. Each step feeds the next. The first three show what this site holds right now for the papers cited here.

1. 1

   Find the paper

   Not done yet

   A paper enters the corpus as a record: DOI, title, authors, journal, date and abstract, taken from Crossref or the publisher.

   • 0 of 12 cited papers with a DOI are in this corpus.
   • Counted in this deployment on Oct 28, 2026.
   • On 8 Oct 2026 the papers cited by the protocol chart were added to the staging corpus as records with an abstract only. No PDF is stored. The counts in these steps are read from the deployment you are on.

   Open the corpus
2. 2

   Get the full text

   Not done yet

   Where the publisher allows it, the PDF is stored and converted to text, tables and equations. Values can also be read from open-access full text without storing the PDF. A paper with neither stays at its abstract: only what the abstract states is available, without its tables or methods.

   • None of the papers cited here is in this corpus yet.
   • On 11 Oct 2026 the open-access full text of some of them was read. Paywalled papers, and open-access ones that a publisher does not serve to a script, still have an abstract only. Fetching the rest is part of the “Kolbe literature corpus” to-do, which needs Max’s list (see Our to-dos).
3. 3

   Extract values with their evidence

   Not done yet

   A language model reads the text and tables and writes every value as one row: parameter, number, unit, the conditions it was measured under, the quote it came from, a confidence and the extractor version.

   • None of the papers cited here is in this corpus yet.
   • On 11 Oct 2026 an AI model extracted values from that full text, each with the sentence it came from. A check confirms that the sentence is in the paper and contains the number. All values are pending review. Extraction with a Kolbe-specific specification is still part of the same to-do.
4. 4

   Harmonise

   Not done yet

   Each row is mapped to a canonical parameter where a rule matches, so “power density” means one thing across papers, and its unit is converted to SI. Conversion is not comparability: the species, the reference electrode and the area or volume basis must match too. Those checks are incomplete, so some rows that cannot be compared still reach the models.

   • Canonical names and units for the Kolbe parameters are the first of our to-dos (see Our to-dos). Until then the extracted values keep each paper’s own names and units, so the predictor cannot use them.

   Open the parameters
5. 5

   Model and predict

   Not done yet

   Models are fitted on harmonised rows only: first the effect of one parameter inside each study, then pooled across studies. A prediction returns a value, a unit, an interval, a confidence and its source. The aim is that a missing input is named, never filled with a default; some MESSAI predictors still fill one today, such as an applied voltage of 0.6 V for an MEC.

   • Nothing is fitted for Kolbe yet: every field of the predictor’s answer is pending (see the question at the top of the page).

   Read how the predictions are made
6. 6

   Check it

   Links listed

   The papers behind the protocol chart are listed below with a link to each publisher. The chart’s steps are steps of the shared protocol graph. Signed-in readers can comment on this page. A listed paper is a pointer for checking, not a check MESSAI has made.

   Go to the papers

Parameter–outcome map

Rows are operating parameters, columns are outcomes. Each cell says how much we know about how that parameter moves that outcome. Most of the map is empty. The predictor’s value lies in filling those cells with effect sizes and intervals, starting with the ones that matter most for design decisions. Select a cell for the evidence and how we would quantify it.

How MESSAI turns this into predictions

MESSAI already runs this pipeline on the bioelectrochemistry corpus. For Kolbe it needs a parameter set, a literature corpus and his run data. The steps run in this order because each one feeds the next.

1. 1

   Within-study effects

   Exists in MESSAI

   For each study, estimate how one parameter moves an outcome while the others are held fixed. This is the cleanest evidence for a trend: differences between labs drop out, though confounders inside a study, such as run order or electrode age, remain. MESSAI fits these as population-level Bayesian effects and shows them in the lab’s sensitivity chart (for example, the within-paper HRT → COD removal effect in the artifact of 3 Sep 2026 rests on 8 papers and 60 observations; it has not converged and its interval includes zero). His paper would be the first Kolbe study encoded this way.

2. 2

   Cross-study pooling

   Exists in MESSAI

   Pool the same effect across Kolbe papers with a hierarchical meta-analysis (MESSAI’s /meta-analysis forest plots). This answers whether his trends hold for other electrodes, solvents and labs, and how much they vary between studies.

3. 3

   Correlation structure between parameters

   Exists in MESSAI

   Some parameters move together: concentration raises conductivity, which lowers cell voltage; chain length and hydrophobicity rise together. MESSAI’s learned dependency graph and knowledge graph record links like these (none yet for Kolbe), so that a predictor would not count one effect twice. Correlation across studies isn’t causation, so causal claims rest on step 1.

4. 4

   Compact laws by symbolic regression

   Exists in MESSAI

   MESSAI already uses symbolic regression (PySR) to search for short equations between parameters. For Kolbe the target is a law such as CE_fuel = f(concentration, mean chain length, current density) that a chemist can read and test, rather than a black box.

5. 5

   Predictor with intervals and drivers

   To build for Kolbe

   A hierarchical model with study as a random effect, using the effects from steps 1–4 as priors and his runs as the core data. His own study would be left out of those priors, so his runs are not counted twice. It would return each prediction with an interval and a ranked list of the parameters driving it. The intervals would still need calibrating on held-out studies: MESSAI’s served predictors do not return calibrated intervals yet.

   outcome ~ f(concentration, composition, j, electrode, cell, pH, T, …) + (1 | study) + residual GP

   To be validated by

   leaving one study out at a time. The real test is predicting another lab’s results, not interpolating within his own runs.

   Planned guardrail

   outside the conditions covered by data, the answer is flagged as an extrapolation; a wider interval alone would not make it reliable

Protocol

The Kolbe methodology as a graph of steps, from preparing the feed to calculating CE and energy per litre. Click a step for its setpoints, readouts, precautions and sources. Left to right is execution order, and every step links to the full /protocols graph. A selected step also shows which rows and columns of the parameter map it feeds.

What this adds beyond the paper

• Generalisation. Do the threshold and the chain-length preference hold for other electrodes, electrolytes and labs? The pooled effects and their spread between studies would answer that.
• Parameters he didn’t vary. Current density, pH, temperature and electrode material would get literature-based estimates with wider intervals, until his runs narrow them.
• Interactions. Does the concentration threshold shift with current density or with composition? Single-factor trends can’t show this; a joint model can.
• A design tool (planned). A parameter combination inside the supported region in, a prediction with interval and drivers out. Run backwards, it would answer “which conditions reach 50 % CE_fuel at the lowest kWh/L for this broth?”
• Where to measure next. The empty cells in the map, ranked by how much each would narrow the predictions that matter, become the next experiments.

Illustration: one of his trends as a what-if (not fitted)

His finding that longer chains react first becomes a one-parameter model of the product mix. Each acid makes radicals in proportion to its concentration and a chain-length preference γ, and radicals pair at random. Fitting γ to his per-acid data would give a product distribution, conditional on coupling, for feeds like the ones he tested; it would not predict conversion, CE, yield, fuel volume or energy. It ignores oxygen evolution, Hofer–Moest, chain shortening and depletion during a batch; the full predictor adds those.

Models

The electrolysis step the predictor is about, and a microbial step that could supply acids upstream. No coupling between the two is modelled, so the pair is an illustration, not a process design. Both are shared MESSAI 3D models; drag to orbit.

Data the predictor needs

Two sources, same columns. His runs give clean within-study effects; the Kolbe literature gives spread across labs. Runs with low efficiency or that failed matter as much as good ones. For literature, MESSAI extracts the same columns with a source snippet for every value.

Papers behind this page

Every paper the protocol chart cites, with what it is cited for and where you can check it. Each row links the publisher and, if this site holds the paper, its corpus record.

Our to-dos (MESSAI)

In priority order. Estimates are rough and assume one person part-time.

1. Kolbe parameter set~2–3 daysNot started

   Canonical names and units for the ten parameters and seven outcomes in the map (proposed: carboxylate_total_concentration, degree_of_neutralisation, anode_current_density, kolbe_ce_fuel, specific_energy_fuel_volume …), sent to Max to check. Fields that get filtered become database columns. Everything else depends on this.

2. Kolbe literature corpus~1–2 weeksNeeds Max’s list

   The 20–50 papers he picks plus a search for the rest, fetched, extracted with a Kolbe-specific spec, harmonized. On 7 Oct 2026 the corpus held one Kolbe paper. By 11 Oct 2026, 54 more had been added to the staging corpus, 47 of them with an abstract, and 434 values had been extracted from the 9 whose open-access full text could be fetched; none is yet harmonised or reviewed. The other 45 have no full text in MESSAI, and which of them are paywalled is not recorded, so no count is given. This is the main source of cross-study evidence. What the deployment you are on holds now is counted under “How it works”.

3. Encode his paper as within-study effects~2 days after dataWaiting on data

   Concentration, composition and cell-design effects from his runs, checked against the trends he reported.

4. Parameter map v1 with numbers~1 week after the two aboveLater

   Replace the evidence states on this page with effect sizes, intervals and the number of studies behind each cell. Add the links to MESSAI’s knowledge graph.

5. Predictor v1~2 weeksLater

   Hierarchical model with study as a random effect, conformal intervals, ranked drivers, and a leave-one-study-out validation report including where it fails.

6. Interactive prediction page~3–4 daysLater

   /collab/pohl in English and German: parameter inputs, prediction with interval, drivers, and the live parameter map.

7. Next-measurement ranking~1 weekLater

   Rank the empty cells by how much measuring them would narrow the predictions Max cares about, and turn the top ones into suggested runs.

8. Supporting fixes~2 days + roadmapNot started

   Product-rate unit harmonization and reference-electrode conversion to SHE; the measurement upload route (roadmap C3-M1, planned 21 Nov) so lab runs land with “measured” provenance.

Deliverables to UFZ

L0 is the day his literature list arrives, T0 the day his run table arrives. Nothing is promised before the data exists.

1. This weekParameter set and run-table template

   The proposed Kolbe parameters and units for him to correct, the CSV template, and the literature request.

2. L0 + 2 weeksKolbe evidence table

   What the literature has measured, under which conditions, with source snippets. Shows at a glance which parameters the field has studied and which it hasn’t.

3. T0 + 1 weekHis paper encoded

   Within-study effects from his runs, reproducing the trends he published, or a note on what doesn’t reconcile.

4. T0 + 4 weeksParameter map with numbers

   Effect size, interval and number of studies for each filled cell; empty cells ranked by importance.

5. T0 + 6 weeksPredictor v1 and validation

   Predictions with intervals and drivers, tested by leaving one study out at a time.

6. T0 + 8 weeksInteractive page, suggested runs, methods note

   The prediction tool in MESSAI, the top suggested measurements, a co-authored methods note draft, and the dataset on Zenodo with a DOI if UFZ agrees.

What we ask of Max and UFZ

• The run table behind the 2026 paper, plus unpublished runs, especially any that vary current density, pH, temperature or electrode, and the failed ones.
• 20–50 key Kolbe and electrobiorefinery papers he trusts, to seed the cross-study evidence.
• A check of the parameter set and of the “known direction” cells in the map.
• 30 minutes every two weeks to review results, and agreement on data terms and authorship.

Open questions for Max

1. Which parameters besides concentration and composition did he vary, in the paper or in unpublished runs? Current density, pH, temperature, electrode?
2. Which cells in the map would change his next experiment if we could predict them?
3. Does he suspect interactions? For example, does the ≈0.1 mol/L threshold move with current density or chain length?
4. Are the “known direction” cells right for his system? Platinum versus carbon, high current density favouring Kolbe.
5. Which other groups’ Kolbe data does he trust, and which would he exclude?
6. Real broth: how do impurities and real titres compare with synthetic mixtures?
7. Would he run experiments the model suggests, as a loop?