Materials Scientists

Google DeepMind's GNoME (Graph Networks for Materials Exploration), published in Nature in November 2023, used graph neural networks to predict the stability of 2.2 million new crystal structures — 380,000 of which were validated as thermodynamically stable, compared to approximately 48,000 known stable materials in prior databases. Microsoft Research's MatterGen (2024) went further, implementing a conditional diffusion model that generates novel crystal structures directly conditioned on target properties such as band gap, symmetry, or bulk modulus — effectively inverting the materials discovery pipeline. These are not research prototypes; GNoME candidates are being actively pursued experimentally by external labs using the published dataset.

A new class of ML models — universal neural network interatomic potentials (UNIPs) — trained on massive DFT datasets (Materials Project: ~150K structures; Alexandria: ~4M structures) now predict atomic forces and energies across nearly the full periodic table with near-DFT accuracy. MACE-MP-0 (University of Cambridge, 2023), SevenNet (Seoul National University, 2024), CHGNet (Berkeley, 2023), and ORB-models (Orbital Materials, 2024, a VC-backed startup) achieve mean absolute errors of 20-50 meV/atom versus DFT — acceptable for most screening applications — while running 1,000-10,000x faster than DFT. These models are open-source and running on consumer GPUs. The AiiDA and Atomate2 workflow managers are integrating UNIP backends, enabling full automation of computational screening campaigns.