MaterialsPilot: An Execution-Feedback Framework for Generative Design of Complex Atomistic Architectures

Generating precise complex atomistic architectures from natural language specifications is a frontier in AI-driven materials discovery, critical for realizing functional systems like catalytic active sites and hetero-interfaces. While Large Language Models (LLMs) offer a powerful interface for such tasks, mapping abstract semantics to rigorous structural constraints remains unsolved as LLMs inherently face two deficits: first, the lack of quantitative physicochemical priors leads to topologically imprecise outputs; and second, open-loop generation without physical verification fails to resolve intermediate spatial conflicts, causing error propagation to invalidate coupled multi-step assemblies. To bridge this gap, we introduce MaterialsPilot, which formulates generation as iterative code optimization. It incorporates two key mechanisms: (1) Hierarchical Retrieval to bridge domain knowledge gaps by providing precise domain schemas; and (2) a Physics-Aware Feedback Loop to master complex logic by autonomously debugging execution and physical violations. Across four complex material modeling tasks and seven diverse LLM backbones, MaterialsPilot increases the rate of fully successful generations from 22.0% to 66.4%, while also producing structures that better satisfy user-specified constraints and physical plausibility checks. These results establish a model-agnostic framework for physically reliable language-driven atomistic design.