Materials ML that respects the physics.
Property prediction only works if the model understands structure-property relationships, not just pattern-matches on composition strings. This page explains how Matforgelab handles that, and where the model's limits are.
Three modeling decisions that set the approach.
Graph-based materials representation
Compositions and crystal structures are encoded as graphs: atoms as nodes, bond length and proximity as edges. This representation preserves lattice symmetry and lets the GNN generalize across crystal systems. Descriptor-based methods require you to enumerate and hand-tune features for each material class. Graph-based encoding does not.
Transfer learning on small experimental datasets
Base models are pre-trained on open DFT-computed property collections, giving the GNN broad coverage of structure-property patterns across material classes. Fine-tuning then adapts the model to your specific system: your alloy composition range, your polymer backbone, your electrolyte salt family. A few hundred measured data points is typically enough to make the fine-tuned model useful. This is how materials informatics closes the simulation-to-experiment gap without requiring you to rebuild from scratch on every new material system.
Bayesian active learning loop
Each round, the acquisition function decides which composition to nominate next by balancing two signals: predicted performance and prediction uncertainty. High-uncertainty candidates in promising regions of composition space are prioritized over grid-search sweeps or random samples. The loop closes when you upload characterization results: the model updates its posterior, and the next nomination batch becomes more targeted. Campaign history shows convergence rate across rounds so you can calibrate how many synthesis iterations a campaign will realistically need.
Open science foundations, proprietary fine-tuning.
Base models are pre-trained on publicly available DFT-computed property databases and published experimental literature. Property domains currently covered: electrochemical (specific capacity, ionic conductivity, oxidation stability), thermomechanical (Tg, tensile modulus, elongation, yield strength), and optical. Your experimental characterization data fine-tunes models within your account and is never pooled with other customers' data.
What Matforgelab does not replace.
Matforgelab narrows composition space and ranks candidates. It does not replace synthesis expertise, characterization judgment, or domain knowledge about processing constraints, safety handling, or scalability. Validated properties still require lab measurement. The model's confidence intervals tell you where predictions are reliable enough to act on and where the uncertainty is too wide to skip synthesis.
Predictions are also bounded by the coverage of the training distribution. If your material system sits far outside the DFT literature and your experimental dataset is small, the model will say so. We do not extrapolate quietly.