Generative Matter V3: An open dataset of AI-generated synthetic materials for architecture

Approximately 120,000 candidate compositions derived from Google DeepMind's GNoME stable-materials corpus, each carrying an AI-drafted workshop recipe, scored and filtered for architectural use, and narrowed to the 245 materials published in the Synthetic Types Atlas (2026). Five layered, documented, machine-readable tables (JSONL + Parquet): materials, recipes, scores, architectural candidates, and the published atlas selection. Generated and curated by AI agents; conceived and directed by Daniel Koehler, State Academy of Art and Design Stuttgart. Supersedes the 2025 Generative-Matter-V2 architectural filter. Note: the data is a set of machine-generated hypotheses, not validated laboratory protocols. Predicted stability is not synthesizability; firing temperatures are best-estimates; safety flags are guidance, not safety data sheets.