Agentic Modeling Framework for Materials‐to‐Device Cross‐Scale Design of Perovskite Solar Cells

ABSTRACT Perovskite solar cells (PSCs) have progressed rapidly over the past decade. As the core functional layer, the intrinsic properties of perovskite materials, including band structure, carrier transport, and optical response, strongly influence PSC performance. Theoretical calculations support absorber screening and device‐performance prediction. However, most studies remain confined to either materials‐level investigation or device‐level evaluation, without a general framework linking microscopic material properties to macroscopic device performance. To address this gap, we develop a large language model (LLM)‐driven, materials‐to‐device cross‐scale agentic framework that establishes a closed‐loop workflow from first‐principles descriptor extraction to solar‐cell performance optimization. A Manager Agent coordinates density functional theory (DFT)‐based materials calculations, drift‐diffusion device simulation, and Bayesian optimization of PSC architectures. Through the model context protocol (MCP), 55 in‐house tools and analysis/simulation programs are encapsulated and orchestrated as callable skills, enabling human‐in‐the‐loop review, cross‐scale parameter transfer, task‐state control, simulation execution, provenance tracking, and error recovery. Across 24 multi‐turn sessions and 60 benchmark turns, the agent achieved a pass rate of 96.7%, significantly outperforming OpenClaw (60.0%). This work provides an extensible methodological route for traceable, auditable, and closed‐loop cross‐scale research on PSCs.