Home▸Materials Science▸Active Learning for Machine-Learned Interatomic Potentials

Active Learning for Machine-Learned Interatomic Potentials

Watch a query-by-committee active-learning loop train a machine-learned interatomic potential: an ensemble of models disagrees most on unfamiliar atomic environments, those get flagged for a simulated DFT calculation, and the ensemble uncertainty collapses around them.

Materials Science3DAdvanced60 FPS📱 Mobile-adapted⇄ 2D version

Universal neural-network potentials like M3GNet, CHGNet and CGCNN replace expensive DFT calculations with a fast graph neural network trained on force and energy labels — but which atomic configurations get labeled matters enormously. This simulator renders a small 3D crystal supercell exploring configuration space under thermal jitter, with a committee of ensemble models predicting the force on every atom. Each atom is colored by how much the committee disagrees (its predictive uncertainty); when that disagreement crosses your acquisition threshold, the environment is queried against a simulated DFT calculation, its label is added to the training set, and uncertainty collapses locally — exactly the query-by-committee active-learning loop used to bootstrap real universal interatomic potentials.

⚙ Under the hood

Train a query-by-committee active-learning loop for a machine-learned interatomic potential: watch an ensemble of models disagree most on unfamiliar atomic environments, trigger simulated DFT labeling, and see the uncertainty collapse.

machine learningmaterials scienceDFTactive learningneural network potentialmolecular dynamics

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install