Upload & map
Read CSV or XLSX, choose regression or classification, and explicitly separate process inputs from outcomes.
Data remains in this browser session.Scientist-in-the-loop active learning
From your last experiment to your next best one.
MADGUI turns small experimental datasets into practical, uncertainty-aware recommendations—without code, installation, or surrendering scientific judgment.
01 CSV + XLSX02 Regression + classification03 Single + multiobjective
Recommended next experiments
Make the next batch count.
Multiobjective utilityPerformance 64% · Cost 36%
CandidateTemperatureTimeScore
01M-041900 °C6 h92
02M-0421,100 °C3 h87
03M-043800 °C9 h81
Accuracy screen passedCalibrated uncertainty · Pareto-aware
Built for real materials work
AlloysCeramicsPolymersBatteriesThin filmsProcess optimisation
A defensible active-learning cycle
Model accuracy is a gate, not a footnote.
MADGUI tests what the data can support, warns when predictive accuracy is weak, and separates cold-start experimental design from model-guided optimisation.
Benchmark honestly
Compare GP, linear/logistic, ElasticNet, Random Forest, and XGBoost with out-of-fold predictions.
MAE · RMSE · R² · accuracy · macro-F1Constrain reality
Declare ranges and discrete resolutions, then configure one to three weighted objectives or a desired class.
Infeasible points never enter the batch.Recommend & test
Inspect calibrated candidate uncertainty, trust warnings and permutation importance; then test and return outcomes.
Uncertainty refreshes with every active-learning cycle.Measured implementation benchmark
A full decision cycle.
Measured in seconds.
Reference first-run computation times for MADGUI's current scientific pipeline—not a reduced demonstration. The benchmark follows representative Q1, median and Q3 workloads for experimental campaigns with fewer than 100 measured samples.
Reference workload
Two objectives. Five model families. One calibrated batch.
CSV parsing · 5 process inputs · 2 regression objectives · 10-fold cross-validation · GP, Linear, ElasticNet, RF-100 and XGBoost-100 · trust screening · calibrated ensemble inference across ≈2,200 feasible candidates · 5 recommendations
Q1 reference study20 measured samples
0.57 sParsing, two independent model benchmarks and a completed-experiment-aware batch.
Trust gate active · space-filling fallbackMedian reference study40 measured samples
1.16 sValidated model ensembles, candidate-specific calibrated ranges and acquisition ranking.
Model-guided recommendationQ3 reference study60 measured samples
1.58 sThe same complete multiobjective workload at a larger small-data campaign size.
Model-guided recommendationIncluded in every time: file parsing, cross-validated model comparison, model selection, trust logic and recommendation generation.
Reference conditions: first-run compute-path timing on an Intel Xeon Platinum 8573C after application load. Browser and device timings vary; page download and human interaction are excluded.
Interpretation: 20, 40 and 60 samples are representative Q1, median and Q3 workload assumptions for sub-100-sample studies—not statistics collected from MADGUI users.
Private by design
Your experimental data stays inside your browser.
MADGUI is designed for both academic research and confidential industrial work. Your selected dataset is used locally to run the scientific pipeline—not uploaded to build a data collection.
This browserLocal working session
01Select
Your CSV or XLSX file
02Parse
Rows, units and constraints
03Fit
Models and cross-validation
04Recommend
Next feasible experiments
No site-storage copy
Uploaded rows, constraints, targets and generated results are not written to MADGUI storage.
Browser parsing
CSV and XLSX files are read by the application in your browser.
Browser model fitting
Benchmarking, inference and recommendation calculations run in your browser.
No experimental-data storage
MADGUI does not place your experimental dataset in the site’s persistent storage.
No confidential retention
MADGUI does not receive or retain a copy of confidential academic or industrial data.
MADGUI Fair-Use Charter
Your data.
Your output.
Your discovery.
MADGUI should accelerate scientific work without acquiring a stake in it. These are five public commitments—not optional enterprise add-ons.
Five non-negotiable commitments- 01
Ownership remains yours
For self-service use, MADGUI claims no ownership in the data you select, the outputs generated for you, or discoveries made independently using the platform.
- 02
No stake in discovery
MADGUI receives no equity, royalty, licence interest, ownership interest or success fee from discoveries made using the platform.
- 03
Your data is never sold
MADGUI does not sell, license, broker or otherwise monetize your experimental data.
- 04
No silent shared-model training
Your data is not used to improve shared models. Any future contribution would require a separate, specific and explicit opt-in; ordinary use is never treated as consent.
- 05
Portable by design
The complete campaign—observations, roles, units, constraints, task and objectives—can be exported as reusable machine-readable CSV. Recommendation batches remain separately exportable.
Ownership boundary: MADGUI neither takes nor reallocates rights. Existing obligations to employers, institutions, collaborators, funders and applicable law remain unchanged. Any formal consultation or collaborative research is governed separately by its applicable institutional agreement.
Plain-language privacy promise
The site delivers the application needed to run MADGUI. The experimental CSV/XLSX file you select is not uploaded to or retained by MADGUI.
Continue to the workspace →MADGUI parity workspace
Validate first. Then recommend.
1 Data2 Task3 Validate4 Recommend
Feasible experimental space
5 process inputs
Input parameterMinimumMaximumResolution ResolutionThe smallest allowed step between experimental settings. Candidates are snapped to this discrete grid.
Particle size (nm)
Volume (mL)
Temperature (°C)
Time (h)
Ramp (°C/min)
Model validation · RF 100 trees · XGBoost 100 rounds, depth 3
10-fold cross-validation Cross-validation strategyWith 10 or fewer complete samples, each sample is left out once for testing. With more than 10, the data are divided into 10 held-out folds.
GP · Linear/Logistic · ElasticNet: Fold-local robust scaling Input scalingGP, Linear/Logistic, and ElasticNet map inputs with declared experimental bounds to 0–1. An input without declared bounds uses only the training fold's median and IQR, with safe fallbacks when the IQR is zero. Held-out samples never set the scale. Random Forest and XGBoost remain unscaled.Compare model families before using any of them for decisions.
Recommendation batch
Candidates will appear after the pipeline check
Benchmark, inspect the accuracy warning, then generate a batch.
Scientific boundary
Model-guided candidates use CV-led ensembles and 90% candidate-specific calibrated uncertainty outputs. Calibration is refreshed whenever experimental outcomes change and drives exploration; its coverage interpretation assumes future experiments remain comparable to the current cycle.
Proven in materials laboratories
Three flagships.
One learning loop.
MADGUI did not appear fully formed. Its present framework is the culmination of more than ten years of experience in materials machine learning and experimental optimisation; the peer-reviewed active-learning and Bayesian-optimisation lineage shown below begins in 2019 and is presented in chronological order.
Flagship applications
Three campaigns selected for their scientific outcome, institutional reach, and continuing research impact. Each card describes the approach and result reported in that study.
01 / 032024 · France × Japan · NIMS
Cu2.125Zn0.875SnS4
Kesterite
thermoelectrics
A recent French–Japanese campaign at NIMS used active learning to coordinate heating, sintering, holding, cooling and pressure. The collaboration continues beyond this result.
+60%thermoelectric performance
11starting samples
4experimental cycles
02 / 032022 · NIMS press release
Direct hot extrusion · Nd–Fe–B
Permanent
magnets
From a dataset normally considered too small for machine learning, the loop proposed fabrication conditions that jointly improved coercivity and remanence in anisotropic hot-deformed magnets.
18initial data points
≈40guided experiments
2magnetic objectives
03 / 032022 · Robotics × AI
Multi-component Li–O2 electrolyte
Autonomous
battery discovery
Robotic experiments and Bayesian optimisation searched a vast formulation space and found a counter-intuitive electrolyte without a component previously considered indispensable.
10M+candidate space
3 mo.discovery campaign
≈2×battery lifetime
Published AL/BO lineage · chronological
Read from left to right as a chronological record of domain expertise accumulating across materials systems. It includes six experimental applications, the MADGUI methods paper and a synthesis-first perspective, reconciled with the current NIMS paper list ↗.
01—08
01 / 082019 · Polymer processing
Resin · curing agent · process
Epoxy adhesive
35.8 ± 1.1MPa joint strength
Active learning connected resin chemistry, mixing ratio and curing temperature to a high-strength preparation from a small experimental dataset.
32 initial formulations · 3 cycles · 5 proposals per cyclePeer-reviewed study↗02 / 082020 · Electrocatalysis
PtPdAu
Mesoporous alloy film
47experiments
Bayesian optimisation uncovered an unexpected high-activity methanol-oxidation composition containing only a small fraction of gold.
Less than 1% of the composition space testedPeer-reviewed study↗03 / 082022 · Permanent magnets
Nd–Fe–B
Direct hot extrusion
380kJ m⁻³ energy product
The loop jointly achieved high coercivity and remanence in a scaled-up anisotropic magnet after three guided cycles.
18 initial samples · 1.7 T coercivity · 1.4 T remanencePeer-reviewed study↗04 / 082022 · Robotic electrochemistry
Li–O2
Rechargeable battery electrolyte
>100stable cycles
Automated experiments and Bayesian optimisation identified a synergistic multi-component electrolyte in a vast formulation space.
0.5 mAh cm⁻² capacity · stable solid–electrolyte interfacePeer-reviewed study↗05 / 082024 · Thermoelectric ceramics
Cu2.125Zn0.875SnS4
Kesterite processing
+60%thermoelectric performance
Four cycles coordinated five sintering parameters and delivered a record maximum zT of 0.44 at 725 K.
11 initial samples · only 0.07% of the process space exploredPeer-reviewed study↗06 / 082025 · Thermoelectric processing
GeTe
Pristine thermoelectric
0.86 → 1.14zT at 700 K
A reproducible two-step route matched multi-day processing without alloying and exposed the process variables governing performance.
24 experiments · 11 process variables · >90% shorter processingPeer-reviewed study↗07 / 082025 · Methods & platform
Multi-Application Design GUI
MADGUI
No codefrom data to next experiment
The published interface made active learning and Bayesian optimisation accessible to experimentalists without programming experience.
CSV/XLSX workflow · prediction · optimisation · experiment designMethods paper↗08 / 082026 · Perspective & roadmap
Executable synthesis protocols
Synthesis-first discovery
P → X → yprotocol → structure → property
A peer-reviewed roadmap reframed executable synthesis protocols as design variables and placed closed-loop optimisation at the centre of experimentally grounded materials discovery.
Machine-readable recipes · experimental feedback · self-driving laboratoriesPeer-reviewed perspective↗What transfers across every campaign
The material changes. The scientific discipline does not.
01
Start from what existsSmall, imperfect experimental datasets are treated as evidence—not as an embarrassment.
02
Respect the laboratoryRanges, resolutions, batch sizes and feasible recipes remain part of the optimisation.
03
Return knowledgeEach cycle seeks a better material and a clearer view of the process variables that control it.
A clear scientific boundary
Free to use.
Built for collaboration.
MADGUI provides its complete standardized active-learning workflow free of charge. It is not a limited trial, and generally valid diagnostics, validation, calibrated uncertainty, safeguards and export capabilities are not withheld.
Complete as a general-purpose instrument does not mean that every materials campaign can be standardized responsibly. Real campaigns may still require application-specific scientific judgment, experimental knowledge and interpretation.
Always availablePermanently free
Permanently free
in MADGUI
The complete standardized instrument—without scientifically weakened tiers.
- 01
CSV and XLSX campaign workflow
- 02
Supported models and cross-validation
- 03
Trust screening and space-filling fallback
- 04
Standard constraints and objectives
- 05
Generic calibrated recommendation generation
- 06
Diagnostics and machine-readable export
- 07
Transparent limitations
- 08
Generic interpretation guidance
Campaign dependentApplication-specific
Application-specific
scientific work
Decisions that cannot be automated or generalized responsibly.
- 01
Scientific problem and objective formulation
- 02
Chemistry-informed variables and descriptors
- 03
Batch effects, censoring and data-quality diagnosis
- 04
Bespoke constraints, costs and safety rules
- 05
Custom models and acquisition strategies
- 06
Multi-fidelity or simulation–experiment integration
- 07
Campaign review and stopping decisions
- 08
Materials interpretation, validation and mechanism analysis
What collaboration means
Application-specific expertise is a separate research contribution—not access to hidden software features. Using the free application never depends on entering a consultation or collaboration.
Software feedback
Report an issue with this web application.
Describe a reproducible software problem—not a scientific campaign or a request to interpret experimental results.
No experimental files attachedYour CSV/XLSX data remains in your browser. Only the information entered in this issue form and the stated technical context are submitted.
MADGUI / Materials Active-learning Decision GUI