One platform, every stage of materials & chemistry R&D
From a sparse set of experiments to a ranked portfolio of lab-ready candidates — each module below works standalone or chains into the full pipeline, across catalysis, formulations, batteries, polymers, alloys and molecular chemistry. Every one of them runs no-code, on your data, with every number carrying an evidence label.
The 5-Stage Discovery Loop
Individual Engine Specifications
Data Synthesis
Turn a handful of experimental runs into a statistically faithful dataset. The Matflow Data Synthesis engine fits the joint distribution of your real experiments — not just per-column shapes, but the non-linear correlations between conditions and outcomes.
- Matflow Neural. A conditional-GAN-style mode for tabular data with mixed continuous/categorical columns. The strongest default for larger, more complex datasets.
- Matflow Hybrid. A GAN-style network combined with a copula transform — often the more stable choice on small samples.
- Matflow Statistical. A fully statistical mode that fits marginals plus a correlation structure directly. Fastest, and the most defensible when rows number in the tens.
Data Ingestion
Turn raw lab artifacts into reviewable datasets: upload CSV, Excel, PDFs and instrument files, auto-extract columns and units, then review and accept rows into a clean dataset ready for modeling.
- Instrument files. Parse instrument output directly into structured tables, no retyping.
- Review workflow. Pending → review → accepted, so only validated rows reach your models.
ELN / Samples
The ELN/LIMS digital thread: register samples, plates, wells, aliquots, vials and barcodes, track inventory and storage locations, and keep a full audit trail from bench to dataset.
- Containers & barcodes. Track tubes, plates, boxes and vials with barcoded locations across freezers and shelves.
- Sample lineage. Every dataset row links back to the physical sample that produced it.
Evaluation
Synthetic data is never trusted by assumption. The evaluation module stress-tests every generated batch before it is allowed to influence a model, with 20+ metrics across five report families:
- Quality report. Per-column shape scores (Distribution Match for continuous columns, Category Match for categorical) and column-pair correlation similarity.
- ML efficacy. An Indistinguishability Score that measures how easily a classifier can tell your real rows from synthetic ones — the harder that is, the more faithful the synthesis.
- Dimensionality reduction. Linear Map and Structure Map projections so divergence between real and synthetic is visible, not just scored.
- Privacy Shield. Record-level distance to the closest real row, an overfitting guard, and an exact-duplicate check — relevant the moment synthetic data leaves the lab.
- Anomaly detection. An unsupervised, threshold-free detector that flags the worst-fitting rows for manual review.
Prediction
The Matflow Predictive Ensemble combines five model families — tree-based and kernel-based regressors — because no single algorithm is reliably best across the small, heterogeneous datasets this platform sees. Everything is cross-validated, single-threaded internally, and explainable:
- Physics-derived features. A transform layer (inverse, log, sqrt, square, plus a logit for bounded percentages) applied per feature shape, latent components, and optional composition descriptors for formula columns.
- Explainability Engine. Global Impact (summary, dependence, waterfall), Local Impact per prediction, and permutation importance — you see why the model believes what it believes.
- Honest uncertainty. Conformal prediction intervals computed from the same cross-validation the metrics use, plus an extrapolation flag on every prediction so guesses look different from supported estimates.
Model Hub
Train, serve and share predictive models with automatic algorithm selection, Gaussian-process modeling with conformal uncertainty, and a Model Zoo of pretrained Foundation Models ready to apply or fine-tune.
- Auto-select + GP. Let the hub pick the best algorithm and get honest conformal prediction intervals.
- Model serving. Deploy a trained model and score new candidates through a live endpoint.
Optimization
Discovery reframed as a constrained search problem. The Matflow Optimizer Core navigates the parameter space of compositions and conditions to propose novel, lab-ready candidates — for materials, electrolytes, formulations and alloys:
- Two search strategies. Matflow Adaptive Search (surrogate-guided) and Matflow Evolutionary Search (a genuine multi-objective genetic algorithm).
- Real Pareto fronts. Candidates ranked by TOPSIS against your objective weights, with Pareto dominance computed directly — you get a trade-off front, not a single "best" answer.
- Constraints enforced during search. Compositional constraints (columns that must sum to 100%, essential for formulation chemistry) and per-feature bounds are built into candidate generation, not filtered out afterward.
TEA
Performance is only half the story — the cost-aware optimizer prices every candidate from raw material composition, equipment sizing, CapEx/OpEx factors and end-of-life & recycling economics, then folds it all into one net-value score.
- Five costing modules. Raw materials, equipment, capital and operating expenditure, and end-of-life & recycling value feed a single net-value metric.
- Standard scale-up rules. Equipment sizing uses the six-tenths power law (cost ∝ (size/base)^0.6), matching standard chemical-industry practice.
- Cost in the loop. Run optimization and TEA together in the pipeline so candidates are ranked by performance, stability, selectivity and economics at the same time.
Life Cycle (LCA)
Quantify environmental footprint cradle-to-gate — carbon emissions, cumulative energy demand and other impact metrics — so candidates are ranked by sustainability as well as performance and cost.
- Impact metrics. Compute carbon footprint and energy demand for candidate routes and processes.
- In the loop. Fold sustainability into optimization alongside cost and performance.
Virtual Screening
Rapid in-silico triage of candidate libraries before they ever reach the lab. The Matflow Cheminformatics Engine ranks candidates by structural similarity to a reference molecule, with an optional SMARTS substructure filter for follow-up design.
- Reference-based ranking. Score every molecule in a library against the chemistry you care about, in seconds.
- Substructure control. Restrict hits to scaffolds that match your synthetic route with SMARTS patterns.
- Swiss-army workflow. Use it standalone or chain it into the full pipeline before optimization.
Design of Experiments (DoE)
The inverse question to prediction: not "what will this formulation do?" but "what should we measure next?" The DoE module generates experiment plans that maximize information per lab run:
- Five design families. Full factorial, Box–Behnken, central composite, Latin Hypercube and Sobol sequences.
- Space-filling by default. Plans are generated to cover the design space evenly, so every lab hour earns its information.
- Directly usable output. Download the plan as a CSV ready for the bench, then feed results straight back into prediction.
Active Learning
Close the loop: model predicts → lab measures → model updates. The Matflow Active Learning Engine proposes the next experiment batch from an existing optimization job's history, prioritizing the points where new measurements will teach the model the most.
- Bayesian core. A Gaussian-process-based approach to expected information gain.
- Graceful fallback. When the full engine is unavailable, a heuristic surrogate runs instead — the response always states which engine actually ran.
- Iterative by design. Each round of results sharpens the next batch of recommendations.
Full Pipeline
Every stage above can run standalone — but the platform really shines when they run end to end. The Pipeline page chains Data Synthesis → Evaluation → Prediction → Optimization (plus TEA, Virtual Screening, DoE and synthetic-data nodes) into a reproducible computational graph.
- Visual graph editor. Connect nodes, configure each stage, and run the whole workflow from one screen.
- Automatic handoffs. Each stage's output becomes the next stage's input — no manual export/import.
- Reproducible runs. Every pipeline run is recorded as a job you can revisit, rerun and share with collaborators.
Reaction Kinetics
Map reaction networks as DAGs with retrosynthesis support, and fit chemical reaction kinetics with mass-transfer diagnostics from experimental data.
- Reaction networks. Build, explore and share reaction-graph / retrosynthesis DAGs.
- Kinetics fitting. Fit rate laws and diagnose mass-transfer limitations.
Atomistic
Structure-aware composition descriptors and candidate screening at the atomistic tier, with an optional MACE machine-learned interatomic-potential tier for higher-fidelity estimates.
- Composition descriptors. Structure-aware features derived straight from formula and geometry.
- MACE MLIP tier. Optional learned interatomic potentials, reported honestly.
Physics Fabric
A layered physics stack spanning quantum mechanics, molecular docking, molecular dynamics and free-energy perturbation — price out interactions at any resolution from one workspace.
- Four tiers. QM, docking, MD and FEP share one evidence-labeled workspace.
- Cost-aware dispatch. Choose the tier that matches your accuracy and compute budget.
Materials & HPC
Query crystal lattices, inspect 3D structures and calculate surface energetics, then dispatch Quantum ESPRESSO DFT jobs to Slurm HPC supercomputers and monitor them in place.
- Materials lookups. Search lattices and structures by composition.
- HPC dispatch. Submit, monitor and cancel DFT jobs on remote clusters.
Materials Explorer
Condition-based search across your own experiment datasets — "every run where temperature ≥ 700, Ni loading 5–20 and conversion ≥ 80%" — returning real rows from your files, per dataset.
- Natural conditions. Stack ≥, >, ≤, <, =, ≠, contains and missing predicates across any column.
- Saveable match sets. Persist the result as a reusable matched set for downstream modeling.
Campaigns
The primary scientific object of the platform: a campaign owns an objective and links every dataset, run, prediction and report that participated, keeping a whole discovery project traceable from first experiment to final decision.
- Lifecycle states. Move candidates from draft through reviewed and approved to sent-to-lab, then record the measured outcome back into the loop.
- Domain packs. Start from pre-built templates for catalysis, formulations, energy and more.
Formulations & CASE
Engineered for coatings, adhesives, sealants, elastomers, paints, and resins. Generates Simplex Lattice and Extreme Vertices mixture designs (sum = 100%), matches resins using 3D Hansen Solubility Parameters (HSP), and computes live Cost and VOC compliance.
- Simplex mixture DoE. Generates lattice, centroid, and constrained extreme vertices matrices respecting exact 100% mass/volume sums.
- HSP sphere matcher. Calculates 3D Hansen distance Ra, Relative Energy Difference (RED), and optimizes green/VOC-exempt solvent blends.
- Real-time Cost & VOC. Computes $/kg, $/L, solids %, Kendall-Monroe log-viscosity, and EPA Method 24 & SCAQMD Rule 1113 compliance.
Formulations & CASE (Depth)
The depth tier for formulations: mixture DoE, Hansen solubility and solvent screening, BOM property bands with interval predictions, and reformulation under target constraints.
- BOM property bands. Interval predictions across a full bill of materials.
- Targeted reformulation. Reformulate to hit target properties and compliance.
Polymers & Plastics
Predict thermomechanical properties directly from monomer / repeat unit SMILES and model multi-component copolymer blends using the Fox and Gordon-Taylor equations.
- Thermomechanical QSPR. Predicts Tg (with Fox-Flory Mw dependence), Tm, melt viscosity / MFI, Young's modulus, tensile strength, and elongation.
- Copolymer & blend modeling. Calculates blend Tg across composition curves and evaluates Flory-Huggins chi_12 miscibility.
- Commercial polymer database. Pre-loaded reference profiles for HDPE, PP, PS, PMMA, PET, PC, PLA, PA6, PA66, PVC, PEEK, PDMS.
Batteries & Energy
Optimize multi-component liquid electrolytes, calculate Vogel-Tammann-Fulcher (VTF) ionic conductivity across temperatures, evaluate electrochemical voltage stability windows, and simulate battery cycle-life degradation.
- Electrolyte optimizer. Blends carbonate solvents (EC, DMC, EMC, DEC, PC, FEC), Lithium salts (LiPF6, LiFSI), and passivating additives (VC, PS).
- Voltage stability window. Calculates HOMO/LUMO redox boundaries vs Li/Li+ to ensure compatibility with high-voltage cathodes (NMC811, LNMO).
- Cycle-life degradation. Physics-informed SEI growth forecaster simulating capacity retention curves Q(n)/Q0 and EOL cycle count.
Batteries & Energy (Depth)
The depth tier for energy storage: cycle-life fitting with remaining-useful-life (RUL) forecasting, electrolyte transport mixing and candidate screening with aging metrics.
- Cycle-life + RUL. Fit capacity fade and project remaining useful life.
- Electrolyte mixing. Screen multi-component transport and compatibility.
Metals & HEA Alloys
Design high-entropy alloys (HEA / MPEA) using Miedema thermodynamic featurization and predict single-phase solid solutions vs intermetallics using Guo and Hume-Rothery criteria.
- Miedema featurization. Calculates VEC, atomic size mismatch delta, Delta H_mix, Delta S_mix, and the Yang-Zhang Omega parameter.
- Phase stability classifier. Predicts FCC, BCC, HCP, dual-phase duplex, complex intermetallics (Laves/B2/sigma), or metallic glass formation.
- Mechanical estimations. Estimates solid-solution hardening yield strength (MPa), Vickers hardness (HV), density, and modulus.
Aging & Stability
Model aging and stability across materials — fit decay kinetics and forecast performance degradation to plan regeneration, replacement or end-of-life handling.
- Aging fits. Model performance loss over time and cycles.
- Lifetime forecasts. Predict when performance drops below threshold.
Reactor Engineering
Size chemical reactors and simulate flowsheets from first principles, matching standard chemical-industry scale-up practice before committing to hardware.
- Sizing calculations. Dimension reactors for target throughput and kinetics.
- Flowsheet sim. Link unit operations and see the whole process balance.
3-Step Auto-Wizard
Built specifically for bench chemists and lab researchers: drag and drop your spreadsheet, let Matflow auto-detect columns and mixture constraints, and receive the Top 5 next recommended experiments with AI reasoning.
- Smart column auto-detection. Instantly classifies ingredients, process conditions, targets, and formulation mixture sums.
- Bayesian surrogate recommender. Active learning engine proposes highest-gain candidate experiments with 95% confidence intervals.
- 1-Click bench protocols. Export recommendations directly into Bench Logger, CSV spreadsheets, and printable lab protocol sheets.
Self-Driving Lab
Closed-loop autonomous experimentation: plan a campaign, auto-generate executable robot protocols, dispatch to automation, and pull measured results back to retrain the surrogate — with safety checks in the loop.
- Protocol generation. Emit validated Python protocols ready for lab robots.
- Results & safety. Collect outcomes and gate runs behind safety validation.
Community Hub
Browse, verify, star and publish versioned models, featurizers, protocols and recipes from the community — every asset carries a SHA-256 checksum so a tampered artifact is always detectable.
- Verified registry. Checksummed, versioned scientific assets you can reuse with confidence.
- Publish & share. Contribute your own featurizers, protocols and recipes back.
Collaboration
LiveDesign-parity collaborative design rooms — presence, review boards, hypothesis sign-off and annotations let the whole team critique candidates in real time.
- Review boards. Sign off hypotheses and annotate candidate sets as a team.
- Live presence. See who is in the room and what they are editing, live.
Dossier Reports
Generate publication-quality dossiers from any campaign — pick HTML, Markdown or PDF and get results, plots and conclusions assembled into one shareable document.
- Multiple formats. HTML, Markdown and PDF export from the same campaign record.
- Audit-ready. Every figure and number carries its evidence label.
Export & Integrations
Move results out and bring tools in — RO-Crate packaging, OPTIMADE endpoints, an MCP server and public data export keep Matflow interoperable with the wider scientific stack.
- Open standards. RO-Crate packaging and OPTIMADE endpoints for materials data.
- MCP & API. Expose your workspace to external agents and tools.
And the tools that tie it together
Analysis, collaboration and automation around the core workflow.
Plotter
Explore datasets and run outputs with a full visualization workbench — scatter, line, bar, distribution and correlation views without writing code.
Launch tool →LiveStudio Collab
A multi-user discovery workspace that pairs run comparison with live collaboration — synthetic vs. real, model vs. model, candidate vs. candidate — so the differences drive your next decision together.
Launch tool →Reports
Turn any run into a shareable, exportable report with the results, plots and conclusions in one document — including public links for collaborators.
Launch tool →AI Copilot
A goal-driven copilot that guides you from objective to campaign, assists with experiment design, and answers platform questions with grounded, cited answers.
Launch tool →PDF knowledge extraction
Pull tabular formulations straight out of literature PDFs and turn them into datasets, instead of retyping numbers from a paper.
Launch tool →Data enrichment
Augment your experiments with computed descriptors and structure-derived features before modeling.
Launch tool →See it on your own data
Every module is free to try on a small dataset. Create an account, upload a CSV, and run the whole pipeline in minutes.