How can we help?

Matflow is a web-based, AI-powered platform that accelerates chemistry and materials discovery and optimization. It turns small, sparse experimental datasets into actionable candidate shortlists. Evidence is shown per result: data-driven models report held-out validation where applicable, while analytic and screening outputs state their engine and evidence class and still require experimental confirmation. It is a no-code tool: you upload tabular experimental data and the platform handles the AI/ML in the background.

Click "Get started" on the home page and register with name, email, password and institution. You can sign in and explore immediately; where card verification is enabled, work creation unlocks after a $0 verification within the grace period shown in billing settings. Administrators can suspend or restrict an account at any time.

Matflow is optimized for tabular data and accepts standard CSV files (and Excel). The first row must be column headers, one experiment per row, and missing values should be empty or use a consistent marker such as NA. Keep headers unique — duplicate columns are a common silent failure.

The platform is designed for the small, sparse datasets common in materials and chemistry research. A good starting rule: fewer than 100 rows suits Matflow Statistical; 100–500 rows suits Hybrid or Statistical; above about 500 rows Neural becomes the strongest default. Always confirm with Evaluation before modeling.

The Documentation section covers every stage in detail, and the Tutorials section walks you through real workflows step by step. The in-app AI Copilot is also grounded in the docs and answers with citations.

Data Synthesis expands a small real dataset into a larger, statistically faithful synthetic one. Evaluation scores how well the synthetic data matches reality. Prediction trains an ML model to predict material or chemical performance from conditions. Optimization searches for new high-performance, low-cost candidates. Run each stage standalone, or chain them end to end on the Full Pipeline page.

Yes — every stage runs standalone. Use the Pipeline page when you want a stage's output to flow automatically into the next.

Rule of thumb from the docs: under 100 rows, try Matflow Statistical first; 100–500 rows, Hybrid or Statistical; over 500 rows, Neural is the strongest default. Each mode learns the joint distribution differently — Statistical fits marginals and correlations directly, Hybrid anchors a GAN with a copula transform, Neural targets larger, multi-modal datasets. Confirm the batch with Evaluation.

Generated batches are never trusted by assumption. Evaluation scores distribution and category match, correlation similarity, classifier indistinguishability, privacy (closest-record distance, duplication, overfitting guard) and anomaly flags on the worst-fitting rows — the full metric list is in the docs and on the Science page.

It means the input lies outside the region your data supports, so the model is guessing rather than interpolating. Treat flagged predictions as hypotheses to test, not numbers to trust.

Each tier defines limits on jobs per hour, concurrent jobs, data points per dataset, synthetic rows per run, maximum upload size and total storage. All plans include the full pipeline — the limits reflect real infrastructure capacity, not artificial caps.

Go to Settings → Billing. Upgrades apply immediately and the new limits are live as soon as payment confirms; you can change or cancel from the same page, with no cancellation fees.

Quota warnings are surfaced in the app before you run out, and requests that exceed the limit are rejected with a message naming the limit that was hit and when it resets. Limits reset hourly for job counts.

Use the "Forgot password" link on the sign-in page. The reset email link is valid for one hour and you can request a new one any time. A successful reset signs out your other sessions; passwords must be at least 8 characters.

Settings → Data & account. Deletion deactivates the account immediately and permanently removes your data after 30 days — during that window, contact support if you change your mind.

Yes. Files are stored scoped to your account and never readable by another user's requests. See the Security page for the full model, and the Privacy policy for what is collected and retained.

No. Every model Matflow trains for you is trained on your data alone, for your account alone, and is not merged into any global or shared model.

Yes — Settings → Data & account lets you export a copy of your profile, dataset list and run history at any time.

Copilot messages are processed by a third-party LLM API to generate responses. Do not paste anything you would not want to leave the platform.

Check the job's detail page for the error message: it states what happened, why, and what to do. The most common causes are incorrect column types, a target with no variance, or a dataset exceeding your tier's row limit.

Confirm the file is CSV or Excel, the first row is a header, and the file is under your tier's size limit. Column headers should be unique — duplicate names are a common silent failure.

Ensemble training is single-threaded internally to avoid CPU oversubscription; a training run on a few hundred rows typically completes in minutes. Very large datasets plus heavy feature engineering is the main slowdown.

The Contact page has a "Bugs / Issues" category, or use the in-app Feedback. Include the job ID if the bug is about a specific run.