Alexandria 2.0: AI-Driven Discovery and Open Data Infrastructure for Materials Design

The Alexandria V2.0 database represents a next-generation platform for AI-assisted materials discovery, integrating generative models, graph neural networks, and universal machine-learning interatomic potentials into a unified open-science framework. In its latest expansion, Alexandria combines large-scale generative structure prediction with active-learning energy refinement to increase the yield of near-stable compounds from 36% to 99% within 100 meV/atom of the convex hull. This effort produced 1.3 million new DFT-calculated materials, including over 74 000 predicted stable phases, expanding Alexandria to 5.8 million total structures. The resulting dataset—comprising both equilibrium and 14 million out-of-equilibrium configurations—serves as a foundation for training universal machine-learning potentials and generative models. By openly releasing all data, models, and workflows under permissive licenses, Alexandria establishes a community-scale infrastructure for accelerating AI-driven materials design. Our analysis highlights emerging correlations between structure diversity, coordination topology, and stability, revealing fundamental patterns that can guide future generative models of matter.