Zero-shot evidence-grounded extraction of blood-brain barrier nanoparticle design parameters with open-weight language models

Data-driven materials design and discovery require high-quality, machine-readable data. However, curating data from unstructured text is a bottleneck. Advances in natural language processing-especially attention-based large language models-enable automated information extraction. Here, we present a zero-shot extraction-synthensis framework that coordinates a two-stage ensemble of smaller, open-weight models ( ⩽ 27B parameters) to extract blood-brain barrier (BBB) nanoparticle design parameters from the literature. The Stage-2 synthesis step enforces fidelity by requiring textual evidence for every extracted value. On a 50-article gold-standard set, the framework outperforms Claude Sonnet 4 on key parameters by F ₁-score: Zeta potential (96.2% vs 87.1%), Particle number (73.7% vs 54.3%), and Particle concentration (40.0% vs 18.6%). Applied at scale, we extract 28-32 BBB-relevant parameters per article across 526 articles. We publicly release the machine-readable dataset to support downstream modeling and benchmarking. Our approach demonstrates an accessible, reproducible path to accelerated data curation for AI-driven materials design.