Search pages, resources, and actions
Discover models, datasets, and services for materials science.
Matbench Discovery — an automated benchmark for ML models on materials stability prediction. Provides standardized evaluation methodology and a public leaderboard comparing UIPs and property predictors on crystal stability classification.
MODNet — Material Optimal Descriptor Network. A neural network designed for small-data materials property prediction. Uses transfer learning and composition-based descriptors to achieve strong performance with limited training data.
MatterSim by Microsoft Research — a deep-learning atomistic model across elements, temperatures, and pressures. Supports energy, force, phonon, and phase-diagram prediction.
ORB v3 by Orbital Materials — a fast, GPU-optimized universal interatomic potential for molecular dynamics. Designed for production use with emphasis on inference speed.
Atomistic Line Graph Neural Network — predicts 52+ material properties by leveraging both atom-bond and bond-angle graphs. Part of NIST JARVIS infrastructure. Top performer on the JARVIS leaderboard for multiple properties.
Crystal Graph Convolutional Neural Network — one of the earliest GNN architectures for crystal property prediction. Directly learns from crystal structure without manual feature engineering. Predicts formation energy, band gap, and stability.
MatErials Graph Network — graph neural network for predicting molecular and crystal properties. Excels at band gap and formation energy prediction. Uses global state, node, and edge features.
MACE-MP-0 — a foundation model for atomistic simulation built on the MACE equivariant message-passing architecture. Trained on MPtrj (>1M structures). 5.6M parameters. State-of-the-art accuracy on energy and force prediction across the periodic table.
Materials 3-body Graph Network — a universal potential trained on ~187K Materials Project relaxations. Supports structure relaxation, molecular dynamics, and property prediction. 225K parameters.
Crystal Hamiltonian Graph Neural Network — a universal interatomic potential (UIP) that directly learns from the Materials Project trajectory dataset. Predicts energies, forces, stresses, and magmoms in a single forward pass. 412K parameters, runs on CPU.
