BigMixSolDB: Extraction of a solubility database in solvent mixtures with an uncertainty-quantified large language model-based pipeline

Abstract

The rapidly expanding body of chemical literature contains physical data scattered across unstructured text and complex tables in thousands of publications. Translating this information into machine-readable formats is essential for training data-driven chemical models. The lack of uncertainty quantification in computationally extracted datasets makes manual extraction and curation still necessary. We here address this challenge and present an uncertainty-quantified data extraction pipeline. Our fully automated pipeline utilizes large language models (LLMs) to extract complex tabular and textual chemical data directly from scientific PDFs. We systematically assess several modern LLMs and document-to-text conversion frameworks. We benchmark the uncertainty in extracted data by comparison of our computationally extracted databases with large, existing databases on solubility that are manually curated. Our pipeline achieves between 93 and 79% of the literature contained data, with near-zero median log deviations for structurally matched rows and residual high-magnitude errors concentrated in a small subset of OCR, table-structure, unit-scaling, or reference-dataset discrepancies. We showcase the utility of our pipeline by deploying it to extract data from over 2,000 articles and generate BigMixSolDB, a comprehensive database on solubility in complex mixtures. BigMixSolDB comprises 280,273 solubility entries spanning single, binary, and ternary solvent systems. Our results demonstrate how integrated LLM-based pipelines can be used for literature data extraction that is comparable with manual curation. We envision that these frameworks can be used for large-scale, accurate extraction of literature data into datasets for machine learning applications