Matflow Benchmark Hub

Measured against reality

Every number on this page was produced by actually running Matflow's production pipeline — never transcribed from a slide. Models scored on held-out experimental data, real physics engines benchmarked against the surrogates they replaced, six gold suites with deterministic pass gates, and standard MatBench tasks on official splits. Download anything and re-run it yourself.

R² 0.842
Matflow Predictive Ensemble
MSE 48.2
vs 151.38 linear regression
88.33%
Synthetic data fidelity
4
Validated systems
51
Published validation rows
0.826
Top novel-candidate score

Benchmark Your Own Dataset

Upload your experimental CSV — chemistry or materials, catalyst or otherwise — to run an instant 3-split validation audit (Random, Extrapolative, Cluster) with AutoML baselines and conformal calibration.

Model leaderboard

Regression quality on the held-out experimental benchmark — the Matflow ensemble balances accuracy and stability best overall.

ModelMSE ↓R² ↑Verdict

Synthetic data fidelity

How faithfully generated rows reproduce the statistical structure of real experiments.

Laboratory case study

Four Ni-hydrotalcite catalysts calcined at 300, 600, 700 and 900 °C, dry reforming of methane, each prediction compared against the measured value from a laboratory campaign. Each output modeled independently.

CatalystCalcination °CCH4 exp %CH4 AI %CH4 errorCO2 exp %CO2 AI %CO2 error

The hydrotalcite and activated-carbon experiments above are published in peer-reviewed journals — ACS Omega 2025 (doi:10.1021/acsomega.5c08293) ↗ and Gas Science and Engineering 2025 (doi:10.1016/j.jgsce.2025.205716) ↗.

External validation datasets

Four independent experimental datasets — sparse synthesis data, extreme small-data, high-throughput rig, and a large operational benchmark.

Novel catalyst candidates

Top ten candidates from the optimization search — model predictions for experimental prioritization, with estimated synthesis conditions.

#CatalystScoreCH4 conv %CO2 conv %H2/COTemp °CNi wt%SA m²/g

Benchmark datasets — download and verify

Every suite dataset is stored locally with its license recorded, so anyone can re-run the numbers without a third-party dependency. The D-Epox-Cat-ML compilation is CC-BY-4.0 (Zenodo 15750387) — please keep the attribution when you reuse it.

CC-BY-4.0 · Zenodo 15750387

D-Epox-Cat-ML — cyclohexene epoxidation literature compilation

207 rows · yield target · used in the benchmark suite above

Bundled with Matflow

Ni-hydrotalcite dry reforming of methane (demo dataset)

51 experimental rows · 3 targets (CH4 conversion, CO2 conversion, H2/CO)

Download .csv ↓

MatBench standard tasks

The official MatBench v0.1 benchmark tasks (materialsproject/matbench, MIT), run with Matflow's composition baseline on the official five-fold splits — so these scores use exactly the same test rows and metrics (MAE / ROC-AUC) as the published leaderboard. Datasets are fetched once from ml.materialsproject.org and cached locally.

Loading MatBench tasks…

Real engines vs. legacy surrogates

What the physics engines actually deliver

Matflow runs the real M3GNet interatomic potential (matgl, BSD-3) and the real PyBaMM continuum battery models (BSD-3) rather than curated tables. The table below compares each engine's relaxed lattice against experiment, alongside the old surrogate table — which shows ~0% error only because it contained the answers; it could not predict anything outside its rows.

Loading engine benchmark… (generated by `python -m benchmarks.engine_benchmark`)

Gold suites — the regression gates

Six versioned fixtures with deterministic pass conditions, evaluated against the production code — not a special evaluation build. DataMap 1.00 · Instrument 1.00 · Agent 1.00 · Sci-Record 1.00 · Atlas (published honestly at its real field recall — the quantified reason the human review queue exists) · base-model E2. Each card links to its full public report; every report is regenerable from the runner named on it.

Every gold suite is a versioned fixture set + runner in backend/benchmarks/ + a CI test that fails the build if a gate regresses (e.g. DataMap accuracy below 1.0, any unsafe input↔output flip, np.trapz reintroduced). Full policy on the methodology page.

Research organizations

Public organizations on the platform — opt-in profiles with a verified badge and member counts. Benchmark participation and published datasets/models land here as org-scoped work ships.

Evaluation & submissions

Signed-in users can run a model evaluation on any standardized benchmark task and submit their own (predictions, targets) pairs to be scored by Matflow against the official task format. The task catalog, scored submissions and standardized leaderboards below are public read-only views.

Run an evaluation

POST /api/benchmarks/evaluate · 5-fold cross-validation on the task dataset

Submit your model to MatBench

POST /api/evaluation/matbench/submit · scored against the official task format

Requires a signed-in session — anonymous submissions are rejected with 401. The scored result lands on the public submissions table below.

MatBench tasks

GET /api/evaluation/matbench/tasks · official task definitions with published model references

Loading MatBench tasks…

Scored submissions

GET /api/evaluation/matbench/submissions · every scored (predictions, targets) pair

No scored submissions yet — the first submitted model appears here.

Standardized leaderboards

GET /api/benchmarks/leaderboards · published suite leaderboards across Matbench, JARVIS and Catalysis

Loading leaderboards…

Download the validation dataset

Run the numbers yourself

The Ni-hydrotalcite dry-reforming dataset (51 experimental rows) that anchors the case study above, downloadable as CSV.

Loading validation data…

Validate your own materials

Bring your experimental data — chemistry or materials — and see Matflow's Data Synthesis → Evaluation → Prediction → Optimization pipeline produce lab-ready candidates.