Transformer Maps 9.13 Million Transition-Metal Clusters, Accelerating Materials Discovery

bravo_abad · x · 2026-08-24

Researchers combined automated first-principles calculations with a transformer-based model to build a Cluster Transformer Encoder Network that predicts atomization energies directly from composition, achieving 40 meV/atom accuracy across 9.13 million compositions (30 d-block metals plus C, N, O, S). A progressive sampling strategy enables efficient exploration of chemical space, revealing systematic links between cluster atomization energies and bulk cohesive energies, signatures of heteronuclear stabilization, and unusually stable noble-metal-doped oxide clusters. This demonstrates an AI-for-science paradigm: use expensive physics to build a reliable surrogate, then explore spaces beyond enumeration. Published in Nature Communications (2026).

Original post →

More from Research

Research channel →