Answers are grounded in this paper and cite their sources.
Computational evaluation demonstrates enhanced information extraction accuracy across materials literature, highlighting effective prompt optimization for automated database curation.
Large language model (LLM)‐based structured knowledge extraction from materials literature is hindered by heterogeneous reporting, schema sensitivity, and prompt dependence. In this study, we present IGA‐POLLM, a gradient‐free prompt‐optimization framework that reformulates prompt engineering as a discrete combinatorial search using six schema‐aligned semantic gene components and an improved genetic algorithm with dynamic crossover/mutation scheduling, tournament selection, and elitist retention. Using corrosion inhibitor literature as a model materials literature system, IGA‐POLLM optimizes prompts for strict JSON‐based extraction of five fields from parsed PDFs. A hierarchical F 1‐based fitness function with hybrid matching aligns optimization with practical database construction. On a dataset of 130 papers, IGA‐POLLM outperforms standard genetic optimization ( F 1 of 0.8571 vs. 0.8208) while reducing runtime (5.6 vs. 8.2 h), generalizes across five publishers (mean F 1 > 0.85), and transfers consistently across five mainstream LLMs. The framework is schema‐driven and extensible to other materials subdomains.
No takes yet. Share an insight, caveat, or question.
Xiang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: