Small experimental data in. Ranked candidates out.
Matflow turns sparse lab data from any chemistry or materials domain into working models, cost-aware candidate rankings and next-experiment recommendations — self-serve from $99/month, with an evidence class and engine card on every number.
The honesty layer is part of the interface: a number the model is guessing at looks different from one with real support behind it.
- MEASURED
- COMPUTED
- PREDICTED
- EXTRACTED
- HYPOTHESIS
- DEMO
- 0.24–0.75
- Out-of-fold R² across the MatBench gold regression tasks
- 0.95
- Leave-family R² on the bundled demo corpus (GBP baseline)
- 4
- Catalytic systems with published lab validation
- 6
- Published gold suites, reproducible from their public reports
- 1.00
- Instrument gold pass rate (16/16 fixture checks)
Why materials R&D is stuck
Three bottlenecks every experimental campaign hits — and the platform is built to break.
A typical campaign — catalysis, formulation, battery or polymer — produces tens of rows, not thousands. A model built on that data is data-starved before it starts, and most tooling ignores it.
Raw materials, equipment scale-up, CapEx and OpEx decide which candidates survive. Economics kill as many materials as kinetics do, yet cost rarely enters the model.
The design space of compositions × conditions is far too large to sweep one variable at a time. Deciding what to measure next deserves the same rigour as the measurement itself.
Four stages, run standalone or end to end
Each stage produces a tracked job in Runs; on the Full Pipeline page the output of one becomes the input of the next.
Expands a small real dataset into a larger, statistically faithful synthetic one. Three modes — Neural, Hybrid and Statistical — and generated rows are clipped to the observed range of each column.
Scores the synthetic batch before you trust it: 20+ metrics across quality, ML efficacy, dimensionality, privacy and anomaly reports, plus projections of real vs synthetic rows.
Trains a five-model ensemble with global and local impact explanations, conformal prediction intervals and an extrapolation flag on every prediction.
Searches compositions and conditions under real constraints, ranks candidates by TOPSIS against your objective weights, and computes Pareto dominance directly.
The 5-Stage Discovery Loop
Optimization Module
Searches multi-dimensional composition spaces under strict physical constraints (100% formulation sum) to uncover non-dominated Pareto frontiers.
You can see where every number comes from
The honesty layer is part of the interface: a number the model is guessing at looks different from one with real support behind it.
- Global & local impact — which inputs moved each prediction, and by how much.
- Conformal intervals — uncertainty bands from the same cross-validation as the metrics.
- Extrapolation flag — a visible warning when a prediction leaves the training envelope.
- Engine card — the engine, its capabilities and its limitations on every result.
Beyond the four stages
Every module speaks the same evidence language. The canonical module catalogue is on /features.
Reads your data, drafts work, navigates pages and runs jobs through the same APIs the UI uses. Reads run free; writes, compute spend and billing wait for an explicit Approve click.
Server-rendered GROMACS, OpenMM and LAMMPS decks; MACE, CHGNet, M3GNet and EMT potentials; docking and free-energy networks; QE/VASP HPC and phonons — each with its engine card.
Parsers for XRD, GC/MS, UV-Vis, FTIR, NMR, rheology and cycler files, ELN/LIMS records with sample lineage, inventory and GHS safety, and an offline bench logger.
Group datasets, runs, candidates and reports under one campaign, with pre-built templates and lifecycle states from draft through sent-to-lab to measured outcome.
Shared design rooms with live presence, candidate review boards, annotations and sign-off chains.
REST endpoints with an OpenAPI spec, API keys issued per integration, and a Python client — rate-limited against the same tier as the web app.
Held-out numbers, reproducible end to end
Engine checks, gold extraction suites and the base-model corpus — drawn from the public benchmark reports. When the live endpoints answer, these cards show the latest published run.
Start where your data is
Deep verticals ship with dedicated workspaces; the rest of the catalogue starts from the same four stages.
Fit capacity fade, forecast remaining useful life with conformal intervals, and screen electrolyte mixtures by transport properties.
Simplex mixture DOE, Hansen solubility screening, BOM property bands and targeted reformulation against your targets.
Create a free account and run the full pipeline on demo data before uploading anything of your own.
Formulation, energy, process conditions, scale-up economics and the bench-to-model loop — the canonical use-case tour.
Frequently asked questions
The questions researchers ask most — the full help centre is at /help.
What is Matflow?
Do I need to be a machine-learning expert?
How much data do I need to start?
Is my proprietary data secure?
Bring your experiments. Leave with candidates.
Matflow is a professional research platform: ingest, model, optimize, dossier — every stage inspectable, with evidence labels on every result and reproducible benchmark reports. Self-serve from $99/month.