Large Language Model Directed Iteration for the Design of Ultra‐Stable Sodium‐Ion Battery

ABSTRACT Integrating artificial intelligence (AI) into materials discovery is hindered by the trade‐off between exploration breadth and exploitation depth, while the lack of embedded domain knowledge and systematic chemical space compression further limits intelligent materials exploration under data scarcity. Herein, we propose a literature‐knowledge‐directed iterative workflow that combines large language models (LLMs) with Bayesian optimization (BO) for the rational electrolyte design for ultra‐stable sodium‐ion batteries (SIBs). LLMs serve as auxiliary tools for knowledge mining, while BO accelerates multi‐objective optimization to obtain electrolyte compositions. The optimized dual‐anion hybrid solvation electrolyte (DHSE) constructs a synergistic solvation structure, which induces the formation of a compact, inorganic‐rich, and mechanically robust cathode electrolyte interphase (CEI). The as‐formed CEI layer suppresses parasitic side reactions and accelerates Na + transport by reducing the desolvation barrier. Remarkably, industrial‐scale Ah‐level hard carbon||NaNi 0.3 Mn 0.3 Fe 0.3 O 2 pouch cells assembled with DHSE deliver a high energy density of 107.7 Wh kg −1 , alongside a superior capacity retention of 94% after 100 cycles at 0.5C. This work not only develops a high‐performance electrolyte for practically applicable SIBs but also provides a general human‐agent collaborative framework for intelligent materials development under data‐limited conditions.