UFZ Leipzig collaboration · Model setups and priors

Live collaboration briefing

The data layer for mechanistic bioelectrochemistry.

Corpus-scale parameters, priors, and model setups — mapped to Dr. Benjamin Korth’s work on the thermodynamics, kinetics, and energy efficiency of electroactive microorganisms.

Prepared for Dr. Benjamin KorthDept. Microbial BiotechnologyPriors artifact · 2026-10-28Nitrogen-cycling use case →

The plan on one page

What comes first, what each step needs, how we test the model, and what has gone wrong so far.

  1. Fix the priors that fail their own checks

    Open

    71 of 171 pooled fits miss the convergence bar, and 33 fraction parameters get intervals below 0 or above 1.

    Needs
    Two model changes: a logit link for fractions (#1302) and a non-centred fit (#1303). No data from UFZ.
    Done when
    No interval leaves its physical range, and power density, current density and coulombic efficiency reach R-hat ≤ 1.01.
  2. Add continuous-BES kinetics

    Needs Benjamin

    The corpus holds no K_S or q_max for continuous biofilms, and pooled reactor data cannot stand in for them.

    Needs
    Benjamin’s continuous-BES runs, with flow, acetate concentration and current.
    Done when
    K_S and q_max each have a value, an interval and n, or a stated reason why the data cannot support them.
  3. Test fine-tune on a real dataset

    Needs Benjamin

    Fine-tune runs, but nobody has checked whether it predicts a new run better than the corpus prior alone.

    Needs
    One dataset from Benjamin, with at least one run held back.
    Done when
    The posterior predicts the held-back run better than the corpus prior does.

What the model should do, and how we test it

Each test copies something a researcher really does. Pass bars are proposals until the partner agrees them.

The model shouldHow we test itIt passes whenToday
Give intervals that mean what they sayScore the 95% interval on 5,269 values from 339 papers kept out of the fit.About 95% of the held-out values fall inside.Passes96.1% (95.5–96.6), 28 Oct 2026. Wide: over 5 decades for areal power density.
Beat the simple guessPredict 3,593 values from 590 papers, frozen before scoring, with each paper kept whole.At least 10% lower error than the median for that class and domain.Fails todayNo model passes (2 Sep 2026).
Improve on the corpus with a lab’s own dataFine-tune on Benjamin’s runs, then predict a run it has not seen.Lower error than the corpus prior alone, and the run sits inside the interval.Needs partner data

Measured without partner data

Two checks we can run on published work alone. Both can be repeated, and the numbers are the same on every page.

CheckResultWhat it leaves open
Are the values we extract really in the paper?96% of 308 checked values from research papers are, across 20 papers (31 Oct 2026).Checked by machine, not yet by a person. 101 values from review papers are other studies’ numbers. 41% of true values reach the model.
Does the model match published runs it never saw?Lab simulator: 7 of 17 within 25%. API and chat: 4 of 17, and 8 got no answer (31 Oct 2026).The simulator got 0 of 5 MFC values within 25%. An MFC request with no COD still answers from its materials alone.

What has gone wrong so far

These rows come from our own sessions and checks, not from partner reports. New reports land here.

What happenedWhat it changesFound in
The served 95% interval described the average, not a new experiment: 0.05–0.11 W/m² against a real spread of 0.00013–44.7 W/m².Every fit now has a predictive interval for a new paper, and the old one is labelled as the interval of the mean.Our check
Fine-tune accepted fits the trainer had marked unconverged, if R-hat was 1.05 or lower.Since 27 Oct 2026 it requires R-hat ≤ 1.01, ESS ≥ 400 and no divergences, and refuses the rest.Our check
33 fraction parameters get predictive intervals outside 0–1. Coulombic efficiency reaches −0.38 to 1.12.Those intervals are withheld. The logit link (#1302) is the fix, and it is still open.Our check

How a failure gets back into the model

  1. Report. Comment on this page, or send us the run.
  2. Reproduce. We turn it into a test with the real input and the expected result.
  3. Fix. We change the model or the pipeline at the cause.
  4. Re-test. The test stays in the set, and the result shows here.

We run these steps by hand. None of it is automated yet.

01 · Model setups you’d recognise

Open any configuration in the lab

Each card lists the physics that governs the configuration. The figures on the cards are typical literature ranges with no source recorded in MESSAI yet: orientation, not data. Click a card to open its detail in the live catalog.

02 · Meta-analysis, sliced by system class

Pooled Bayesian priors as forest plots

Pooled + per-system-class posteriors of the population mean with 95% credible intervals (where a typical study lands, not the range of single reactors) — the per-class split (MES / MEC / MFC) is the story. An amber tier label marks a stratum whose MCMC fit hasn’t converged: flagged, never hidden. Open any card in the live explorer.

Coulombic efficiency

Open in explorer →

Forest plot compares MES vs MEC vs MFC — the per-class split is the story.

Stratum
0.1050.4460.788
n

Areal current density

Open in explorer →

Log-scale prior; heavy-tailed spread reflects real cross-corpus variance.

Stratum
0.1703.156.2
n

Applied potential

Open in explorer →

Energy input distribution for MEC — the energetics lever.

Stratum
-0.338-0.0470.244
n

Remediation performance across cathodic and mixed systems.

Stratum
0.4500.6910.932
n

03 · Priors you can drop into your models

Download the posteriors as sampling specs

The export ships each fit as a ready-to-sample distribution (natural-log lognormal for heavy-tailed quantities, normal for linear) with n, units, provenance, and caveats — straight into PyMC / Stan / scipy.

ParameterClassnCentral95% CIConvergedFlags
coulombic_efficiencypooled23037.2%mean33.2% – 41.1%nofailed diagnosticssampler did not converge (converged=false, rhat_max=1.0202, ess_bulk_min=301)
MEC3656.4%mean48.2% – 65%yestrustworthy
MES2370.1%mean61% – 78.8%yestrustworthy
MFC12424.1%mean19.7% – 28.5%nofailed diagnosticssampler did not converge (converged=false, rhat_max=1.04, ess_bulk_min=94)
current_density_arealpooledlog1760.489 A/m²geo. mean0.319 A/m² – 0.756 A/m²nofailed diagnosticssampler did not converge (converged=false, rhat_max=1.0241, ess_bulk_min=283)
MEClog191.6 A/m²geo. mean0.617 A/m² – 4.03 A/m²yesusablemoderate support (n_papers=19)
MESlog137.54 A/m²geo. mean0.99 A/m² – 56.2 A/m²yesfailed diagnosticsLOO unreliable (high Pareto-k) — influential outliers dominate the fit
MFClog1060.287 A/m²geo. mean0.17 A/m² – 0.46 A/m²nofailed diagnosticssampler did not converge (converged=false, rhat_max=1.0147, ess_bulk_min=435)
cod_removalpooled33370.5%mean68.4% – 72.7%yestrustworthy
MEC4467.6%mean62.6% – 72.8%yestrustworthy
MES973.3%mean55.4% – 93.2%nofailed diagnosticssampler did not converge (converged=false, rhat_max=1.0768, ess_bulk_min=37)
MFC22571.5%mean69.2% – 74%yestrustworthy
applied_potentialpooled10-0.0504 Vmean-0.338 V – 0.244 Vyesfailed diagnosticsLOO unreliable (high Pareto-k) — influential outliers dominate the fit

Every row: hierarchical-priors (v2-sota-2026-05-15); n is papers per fit.

Central = the posterior centre of the population mean (a geometric mean for log fits), the same quantity as the 95 % interval beside it. Grades come from the gate the fine-tune endpoint uses: unusable / failed diagnostics = within-study variance collapsed, the trainer did not mark the fit converged (it requires R-hat ≤ 1.01, ESS ≥ 400 and no divergences), unreliable LOO or ν < 1, interval struck through; sparse = fewer than 8 papers; usable = passes the gate with fewer than 20 papers or an unconfirmed LOO.

04 · Continuous-BES acetate kinetics

The acetate-uptake picture your models formalise

Acetate uptake by continuous Geobacter biofilms is commonly modelled with Monod kinetics, q = q_max·S/(K_S+S). The table shows corpus fits for related observable quantities, graded by the same gate as the priors above; several fail it. None of them is K_S or q_max, and each pools every reactor type in the corpus rather than continuous biofilms. That gap is where your continuous-BES data would come in.

Monod saturation (example data)
acetate concentration S →K_S = ?q_max

A shape, not a fit: K_S and q_max are unknown, so neither axis has a scale. The corpus’s pooled areal current density (posterior centre 0.489 A/m²) is a centre across all reactor types, not the q_max of any biofilm, and is not drawn here.

Kinetic quantitynCentral95% CIFlags
Substrate concentrationlog93685 mg/L392 mg/L – 1.18e+3 mg/Ltrustworthy
Acetate concentrationlog63926 mg/L604 mg/L – 1.45e+3 mg/Ltrustworthy
Areal current densitylog1760.489 A/m²0.319 A/m² – 0.756 A/m²failed diagnostics
Peak current1170.00134 A0.000298 A – 0.00273 Aunusable
Exchange current density (j₀)log90.162 A/m²0.0261 A/m² – 1.05 A/m²failed diagnostics
Specific growth rate (µ)100.162 1/h0.0505 1/h – 0.276 1/husable
Biofilm thicknesslog310.0000292 m0.0000115 m – 0.0000735 mfailed diagnostics
Flow ratelog480.322 L/h0.183 L/h – 0.449 L/hunusable

What the corpus doesn’t yet fit — your continuous-BES data would seed these

  • K_S — Half-saturation constant: the acetate-affinity term in Monod uptake q = q_max·S/(K_S+S) — not extracted corpus-wide
  • q_max — Max specific uptake rate: the saturation plateau of acetate uptake — not extracted corpus-wide

05 · Fine-tune with your own data

Turn the corpus prior into your posterior

Paste your own measurements and pick a base stratum. MESSAI runs a closed-form partial-pooling update — corpus prior ⊕ your data → your posterior — in the same sampling-spec format you download above. It reports the shrinkage weight w (how much is borrowed from the corpus). It refuses a base fit whose within-study variance collapsed, whose LOO is unreliable, whose ν is below 1, or that the trainer did not mark converged: it requires R-hat at most 1.01, ESS at least 400 and no divergences, and a fit with no convergence flag is refused too. Until 27 Oct 2026 it accepted fits the trainer marked unconverged when R-hat was at most 1.05; it no longer does. Bases with fewer than 8 papers are accepted with a warning. The numbers you paste are sent to MESSAI’s API for the calculation and are not stored.

Parameter
Base stratum

Struck-through strata are refused by the fine-tune endpoint: the trainer did not mark the fit converged (R-hat ≤ 1.01, ESS ≥ 400), LOO is unreliable or ν is below 1, or within-study variance collapsed so any data you supply would crush the prior. Hover for the specific reason.

Your measurements (Areal current density, original units)
Corpus weight κ1.00

Open decision

Which model should answer lab-design chat?

Lab-design answers take about 50 seconds today. A smaller model answers the same questions in about 19 seconds, for about 3% of the cost, but looks up the corpus less often. We would like your view on which matters more for your work.

Median time per answer
Claude Haiku 5.519 s
Claude Sonnet 4.6Current50 s

looking up papers and datawriting the answer

MeasureClaude Haiku 5.5Claude Sonnet 4.6
Time per answer (median)19 s50 s
First visible activity2.2 s2.4 s
Writing the final answer10.6 s33.7 s
Corpus look-ups per answer47
Answer length (tokens)3,0262,346
Estimated cost per answer$0.007$0.21

What the answers looked like

  • Both

    Both models cited papers from the corpus and said plainly when the evidence was thin.

  • Sonnet 4.6

    Split power density by the area it is normalised to (anode, cathode, membrane) and warned against comparing across bases.

  • Haiku 5.5

    Gave a direct estimate (100–150 mW/m², anode area) from a few acetate papers and flagged it as preliminary.

  • Sonnet 4.6

    Added a section on mechanisms that was general knowledge rather than drawn from the corpus.

Read for 2 of the 10 question pairs only.

The options

  1. 1Keep Sonnet 4.6Most careful with units and bases. Answers stay at about 50 seconds.
  2. 2Haiku 5.5 for lab-design onlyLab answers in about 19 seconds; general questions keep Sonnet.
  3. 3Haiku 5.5 everywhereFastest and cheapest. Every answer relies on a model that looks up the corpus less often.

Which would you pick, and is there a question you would want us to test both models on? Reply in the comments on this page or at founders@messai.io.

Method: 10 lab-design questions per model, sent in alternating order on Oct 26, 2026; no answer fell back to another provider. Times are measured on the server and exclude your network. Prices are the AI Gateway list prices; cost per answer is an estimate from median token counts with 81% of input read from cache.

Where the data is soft — stated up front

  • Extraction is noisy — cross-corpus scatter is large (power density CoV ≈ 10³%); every value ships with n, units, and range.
  • Coverage is partial — a substantial share of indexed papers are not yet joined to extracted values, so distributions reflect the extracted subset.
  • Sparse or non-converged fits are flagged, never hidden.

Priors served from the per-class hierarchical v2 artifact (PyMC NUTS, Student-t). Model setups and meta-analysis link to the live lab and meta-analysis surfaces. Prepared as a collaboration briefing for Dr. Benjamin Korth, Helmholtz Centre for Environmental Research — UFZ.