🤖 AI Summary
Large language models struggle to directly generate atomic structures satisfying global physical constraints such as periodicity and bond lengths, as conventional token-by-token generation lacks physical plausibility. This work proposes ATLAS, a framework that leverages component algebra to translate natural language into verifiable structure-building scripts. By introducing a checklist-based verification scoring mechanism and a JSON-formatted construction engine, the framework enables deterministic replay, rendering the LLM translation process both measurable and controllable. ATLAS successfully demonstrates the generation of fourteen structural classes spanning crystalline to amorphous networks, and can be directly applied to constructing training datasets for machine learning force fields.
📝 Abstract
The automated generation of three-dimensional atomic structures from natural-language descriptions remains a persistent challenge, as large language models lack intrinsic geometric reasoning. We introduce ATLAS, a holistic framework that integrates seven modular components-ranging from a density-based structure analyzer and a JSON-based semantic bridge to a deterministic build engine, physical validator, and intent-aware scorer-to translate user prompts into physically valid coordinates. The build engine executes a component algebra that encompasses named structures, unary and binary operations (supercell, resizing, rotation, translation, union, subtraction, intersection), and geometric variables, while the validator enforces periodic-boundary conditions and system-specific checks on bonding, coordination, vacuum, and composition. ATLAS supports nine major build categories, including heterojunctions, surfaces, nanoparticles, and defects, and is equipped with a score-driven automatic refinement loop that leverages a curated test suite for systematic component-level improvement.