物理条件を守るAI探索で密度汎関数を改良

We demonstrate physics-constrained agentic development of a meta-generalized gradient approximation (meta-GGA) functional using a large language model (LLM) to assist the search and optimization of a band-gap-oriented revision of r2SCAN. No real bonded systems were fitted, preserving r2SCAN's nonempirical philosophy. Across the finalist set, band gaps and several molecular subsets improve relative to r2SCAN; the top finalist, r2SCAN+, reduces the band-gap MAE on a benchmark comprising 24 solids from 1.26 to 0.96 eV and the aggregate MAE on 329 molecular properties from 4.79 to 4.42 kcal/mol. We first curated 78 exchange and 90 correlation candidate correction terms from r2SCAN's dimensionless ingredients, spanning polynomial terms through third degree, exponentials, exponentially damped products, and ratios. Allowing each candidate to combine one to three correction terms from the exchange catalog, the correlation catalog, or both yields about 8 x 10^5 distinct forms, making exhaustive high-throughput screening impractical. We defined the search criteria for the LLM agent using r2SCAN's exact constraints, physical norms, and the targeted iso-orbital derivative response. The agent then combined these criteria with its pretrained knowledge and accumulated search feedback to propose and refine sparse forms, prioritizing terms tied to the iso-orbital response; a second LLM critic screened proposals before deterministic verification. Compared with uniform random search, the workflow learned from prior evaluations, incurred far fewer downstream rejections (0.6% versus 24.6%), and located stronger high-response candidates: 51 agentic candidates exceeded the best random-search response of 1.263, with the overall best reaching 1.331. These results show that agentic search can support density-functional development when flexible hypothesis generation is coupled to automated physical verification.