Data-driven insights into stainless steel pitting corrosion through large-scale literature mining and database construction

Abstract

Pitting corrosion is one of the most prevalent and detrimental forms of localized corrosion in stainless steels exposed to chloride-containing environments. Its initiation and propagation are governed by the combined effects of alloy composition, environmental conditions, and experimental parameters. Although a substantial body of experimental data has been accumulated through decades of research, these data remain largely dispersed across the scientific literature, limiting their accessibility for large-scale statistical analyses and data-driven investigations. To address this challenge, this study proposes a large language model (LLM)-based framework for structured data extraction and database construction from scientific literature. By decomposing the complex information extraction process into multiple subtasks, including experimental condition extraction, pitting potential extraction, and data alignment, the framework enables the automated acquisition and integration of experimental data from unstructured scientific documents. Using the proposed framework, 12,792 candidate publications retrieved using a broad stainless-steel pitting corrosion search strategy were processed, resulting in a structured extraction output containing 36,095 experimental records. The database encompasses key information, including alloy compositions, environmental parameters, and pitting potentials. Stainless-steel-specific descriptive and statistical analyses were subsequently conducted using material-screened subsets of the extracted records. Data quality evaluation demonstrates that the proposed framework achieves strong aggregate performance in structured information extraction from complex scientific texts, while stage-specific validation indicates that pitting-potential extraction and normalization remain more challenging than the overall field-level extraction task. Statistical analysis of the resulting database reveals that existing studies are predominantly focused on room-temperature, neutral chloride environments, as well as austenitic and duplex stainless steel systems. Furthermore, partial correlation analysis revealed conditional associations between the investigated variables and pitting potential. Among these variables, molybdenum exhibited the strongest positive conditional association, whereas temperature showed a negative conditional association. Chloride ion concentration showed only a weak negative association. These trends were generally consistent with established corrosion knowledge and therefore provided a scientific plausibility check for the statistical patterns observed in the constructed database. This work highlights the potential of large language models for knowledge extraction and automated database construction from corrosion-related scientific literature. The proposed framework provides a reusable data resource and technical pathway for corrosion informatics research, while establishing a foundation for future data-driven studies on corrosion mechanism discovery, performance prediction, and materials design.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗