Platform overview · datasheet

From the literature to reactor estimates you can check.

MESSAI turns a curated corpus of microbial-electrochemistry research into a working model of the field — semantic discovery over the papers, a causal map of the parameters, and a modelling stack that estimates reactor performance with stated uncertainty and traceable sources, and withholds an estimate when the evidence is thin.

Domain
MFC · MEC · MDC · MES
Audience
Academic MES researchers
Snapshot
2026-07-06
Architecture
Web platform · Postgres with vector search

What it is

A research platform for bioelectrochemical reactors

Microbial electrochemical systems are reactors where bacteria oxidise organic matter and pass electrons to an electrode — producing electricity (MFC), hydrogen (MEC), desalination (MDC), or targeted chemistry (MES). The field is data-rich but fragmented across tens of thousands of papers with inconsistent units and reporting. MESSAI unifies that literature into one queryable, model-backed system so a researcher can ask what is known, design a reactor, and get a defensible performance estimate — without leaving a single surface.

The foundation

The corpus, in numbers

Live where a simple count allows; documented and dated where a figure is gated by harmonisation. Re-verify against production before external citation.

19,380Research papersThe literature everything is built on● live · Prisma
10,993Full-text PDFsContent-addressed store, ~20 GB● live · Prisma
219,134Extracted measurementsEach carries its DOI; about half carry a snippet● live · Prisma
~25,700Modelable rowsapplicable_to_modeling = true○ measured 2026-07-06
835Parameter definitionsCanonical variable definitions● live · Prisma
2,819Causal DAG edgesMechanistic + fitted dependencies● live · Prisma

Scientific-integrity note. Extraction is imperfect and coverage is uneven — about 40% of papers still lack parameter extractions (ingest outran extraction), and dispersion is enormous (power-density CoV ≈ 1,285%). Point estimates are meaningless without their context, so every number on the platform is designed to ship with n, units, range, and a confidence flag. The binding constraint on modelability is harmonisation and verification, not raw extraction volume.

What researchers can do here

One platform, from paper to checkable estimate

Every surface reads from the same curated corpus and the same prior stack — so what you discover, analyze, and estimate all trace back to source.

How it works

From a raw PDF to an estimate with a measured interval

The same pipeline feeds every surface. Each stage is auditable — you can trace any number on the platform back through calibration, canonicalization, and extraction to the paper it came from.

  1. 1

    Corpus

    19,000+ MES papers in the corpus, with full-text PDFs content-addressed and embedded for semantic retrieval.

  2. 2

    Extraction

    Two-tier LLM extraction pulls measurements with their conditions. Every value keeps its DOI, and about half keep a source snippet.

  3. 3

    Canonicalization

    Raw values are normalized to SI and mapped to canonical parameters, so numbers from different papers become comparable.

  4. 4

    Priors & calibration

    Hierarchical Bayesian priors (PyMC, log-normal) per system class, wrapped in conformal intervals with measured coverage.

  5. 5

    Prediction

    Per-physics-family predictors return every output as {value, unit, ci_low, ci_high, confidence, source}. Interval fields are filled where a calibrated interval exists.

  6. 6

    3D lab

    Design a reactor and see the prediction, the empirical evidence behind it, and the mechanistic loss decomposition side by side.

The prediction contract

Served predictions carry an interval where one is calibrated

Predictions are never single numbers. Each value ships bounds, a confidence, and a source, and is routed by physics family rather than squeezed into a three-type enum.

// every numeric output
{
  value:      0.42,
  unit:       "W/m^2",
  ci_low:     0.19,
  ci_high:    0.71,
  confidence: "medium",
  source:     "hierarchical_prior + conformal"
}

// interval coverage at the 95% target (priors v2, 2026-10-28)
coverage: 96.1%   (5,269 values, 339 papers)
anodic_oxidationMFC · MES-anode
cathodic_reductionMEC · MBES
ion_transportMDC · MNRC
selective_reductionMERC · MEFS
sensorMBES (discrete outputs)

System taxonomy

17 primary types, not three

The corpus is classified against a canonical taxonomy of seventeen primary system types (plus multi-axis subtypes and combination types). Restricting to MFC/MEC/MDC would silently coerce the rest into the wrong shape.

MFCMECMESMDCMSCMEFSMESNORKMERCMNRCMMRCMRBMBESMRECMCDIREVIEWFUNDAMENTALOTHER

An honest inventory

What’s wired vs. what’s dark

Being candid about maturity is a platform value. These are shipped, user-facing surfaces versus capabilities that exist in code but have no live interface yet.

Live & wired

  • Research search + grounded chat — pgvector retrieval, ~25 corpus/ML tools
  • Parameters — catalog, causal DAG, knowledge graph, sweep
  • 3D lab + ML predictions — intervals on served predictions where one is calibrated
  • Hunter — physics violations, anomalies, outliers
  • Admin — corpus stats, extraction processing

Dark (built, no UI)

  • Manuscript-review engine — grounding tools complete, no surface
  • On-demand figure digitization — runs nightly, not interactive
  • Live Bayesian-network design recs — pre-computed, no custom endpoint
  • Research-agent orchestrator — one-shot synchronous coordinator; no durable workflow

Open by default

Nine open packages (one still scaffolding)

The data and tooling that power MESSAI are released as standalone, citable packages — each with a schema, a license, and a CITATION.cff. Use them in your own analysis; they don’t depend on the platform.

New to the packages? Start with the manual or read how the parameter package is consumed downstream.

Under the hood

Four zones, one repository

An Nx monorepo deploys as four independent zones behind a single URL. Writes route exclusively through the API zone; the others read the shared Postgres + pgvector store, with Python ML services for embeddings, OCR, and Gaussian-process inference.

apps/webNext.js · webpack

Research, parameters, hunter, admin, and the homepage.

apps/labNext.js · R3F

The 3D reactor simulator and model catalog.

apps/apiNext.js · Turbopack

All API routes; the only zone that writes to the database.

apps/siteAstro 5

Marketing, learning, and reference content.

Get started

Bring a research question.
Leave with the evidence, its conditions and a calibrated range where one is supported.

Search the corpus, design a reactor, and read literature-based ranges that name their sources, their conditions and where no calibrated range exists.