The whole materials R&D loop, on one platform.

Pull data out of papers and experiments, standardize it against a domain ontology, train models without code, then work back from target properties to the conditions that meet them.

Request a demo See how it works
D3Square knowledge graph: about 900 entities extracted from papers and experiments, colored by type and clustered by relation

What changes when research data lives in one place

Most labs already have the data they need. It is spread across spreadsheets, inboxes, and PDFs, named differently by every group that touched it.

TodayWith D3Square
Experiment data Spreadsheets on personal PCs, email attachments, and systems that do not talk to each other. Experiments, literature, and analysis records in one place, with version history.
Published results Numbers stay in PDF tables and figures that nobody has time to transcribe. A vision model reads tables and figures, and every value links back to its source page.
Terminology Three groups name the same property three ways and write units their own way. Ontology concepts merge the variants and convert units to SI before anything is trained.
Modelling Specialists write code in a separate environment, and the results stay with them. No-code training, with a plain-language report the whole team can read.
Finding conditions Run the experiment again, and again, until something works. Start from the target properties and let inverse design propose the conditions.

Four stages, one loop

Each stage hands its output to the next, and the last one feeds the first: the design stage tells you which experiment to run next.

  1. Collect

    Papers, experiments, lab setup, and inventory land in versioned buckets.

  2. Standardize

    Terms map to ontology concepts, units convert to SI, out-of-spec values are held back.

  3. Learn

    Profile the data, train models without code, and read what they learned.

  4. Design

    Set target properties and search back to the composition and process that meet them.

Active learning picks the experiments worth running next, and the loop starts again.

An AI assistant runs alongside every stage. It knows which screen you are on, answers from your own data and the knowledge graph, and can run platform actions for you.

What each stage does

Collect

Literature extraction

Add papers by PDF, DOI, or PubMed ID, or subscribe to a journal and let new articles arrive on their own. Table and figure pages are rendered at 200 DPI so a vision model reads the numbers a text parser would mangle.

  • Crossref and PubMed search, scheduled journal feeds
  • Every value keeps its source page for one-click checking
  • Extracted values map straight into a bucket column
PDF Vision model, 200 DPI Table 2, p. 7
SampleSinter Tσ (mS/cm)
S-01850 °C12.4
S-02900 °C18.7
S-03950 °C15.2

Mapped to bucket column conductivity

Lab records

Lab setup, inventory, and experiment designs are recorded next to the literature, then grouped into buckets the team can share and train on.

Experiment E-24 Sintering Furnace B, 900 °C
InputLotUsed
Li₂CO₃L-041212.0 g
La₂O₃L-038830.5 g
ZrO₂L-04159.8 g

Added to bucket LLZO sintering

Standardize

Ontology

Two labs measure the same property and the data still will not merge. Ontology concepts pull the variants together, convert units to SI, and hold back out-of-spec values before they reach a training set.

  • Concepts with synonyms, units, definitions, and hierarchy
  • New concepts and rules go through review
  • TTL import and export, EMMO snapshot diffs
ionic conductivity σ_ion 이온전도도 IonicConductivity ontology concept S/cm QUDT unit range rule held back Sample S-02 → hasProperty → IonicConductivity inferred by a transitive rule, source relations kept

Knowledge graph

Entities and relations pulled out of your literature land in a graph you can look at. Filter by type, follow a relation, ask the assistant about it, or send the selection to a bucket for training.

  • Materials, properties, methods, journals, and authors as typed entities
  • Answers grounded in the graph, with the nodes they came from
  • Results respect data permissions: no access, no evidence
Knowledge graph centred on ionic conductivity: solid electrolytes such as Li6PS5Cl, Li3PS4, and Li10GeP2S12, separators, and measurement records extracted from papers

Learn

Data buckets

Upload experimental datasets and they are checked on the way in. Missing values, outliers, and categorical columns are handled in one preprocessing step.

  • Missing values and outliers flagged on upload
  • Statistics, correlation, and SHAP feature importance
  • Version history you can restore
Temp (°C)PressureYieldDensity
1,2403.294.17.85
1,180—91.77.82
1,3103.596.37.91
1,2753.395.07.88

1 missing value flagged in Pressure

Model training

Train straight from a bucket without writing code, compare runs, then read what the model learned. The platform writes the interpretation out in plain language, cautions included.

  • GPR, XGBoost, linear models, and deep learning
  • Cross-validated runs compared on R², MAE, and RMSE
  • Publish a model for prediction and inverse design
AI interpretation Target: yield strength (MPa)
Summary

R² is 0.55: the model explains about half the variance. Moderate, but the patterns are real.

Key findings

MAE 12.1 MPa, RMSE 17.0 MPa, so a few predictions miss by a lot. Cu drives the prediction most, then casting temperature.

Caution

Predictions run high: 104–179 MPa against actual values of 74–164 MPa. Add data before relying on it.

Design

Inverse design

Define the target properties first and let published models work back to the composition and process conditions that meet them: import a model, map its roles, set the design variables, pick a search, run.

  • Single and multi-objective runs on published models
  • Genetic algorithm, Bayesian optimization, grid and random search
  • A Pareto front of candidates worth making, within your constraints
Inverse design result for an aluminium alloy: 2,000 candidate compositions plotted by predicted elongation and yield strength, with the Pareto front highlighted
Aluminium alloy, yield strength against elongation for 2,000 candidates. The 17 on the Pareto front (orange) run from 547 MPa at 13.6% to 428 MPa at 15.2%.

Active learning

When the model is unsure, it says where. The platform recommends the samples that would teach it the most, so each round of experiments costs less.

  • Expected improvement, UCB, and Thompson sampling
  • Each round retrains on what you measured
Train model Find uncertainty Suggest samples Run experiment

AI assistant

The assistant knows which page you are on. On a training run it offers to compare performance, on the knowledge graph it answers from the graph, and on the dashboard it tells you what moved. Ask it to do the work and it operates the platform for you.

  • Vector search over documents plus graph traversal
  • Runs buckets, training, models, inverse design, CAE, and DOE
  • Questions in any language, answers scoped to what your account can see
Dashboard

Morning. Want today’s research briefing?

Training

Two runs finished. Compare their performance?

Knowledge graph

Three papers link this method to conductivity. Show them?

Answers cite the documents and graph nodes they came from.

Built for materials R&D, not adapted to it

Experiments, inventory, literature, properties, and process conditions are objects with the right fields and units, not anonymous columns in a spreadsheet.

Knowledge that carries over

Structured literature and published models become the starting point for the next project instead of one-off files.

Permission-scoped data

Data is shared on purpose. What a user cannot open is also left out of graph queries and assistant answers.

API and LIMS integration

Connect the systems your lab already runs, so data keeps arriving without anyone re-typing it.

Suited to alloy composition and heat treatment, catalysts and reaction conditions, battery and energy materials, and quality prediction in manufacturing.

See it on your own data

Tell us what you are working on, and we will set up a demo around papers and data from your field. We also answer questions about deployment and technical fit.

Request a demo

or write to support@virtuallab.co.kr