Introduction
Samarium (Sm) and Neodymium (Nd) isotopes serve as critical indicators for geoscientists tracking the evolution of the Earth's crust over billions of years. Despite their importance, a vast amount of this data remains "trapped" within the static tables of thousands of scientific publications, primarily in PDF format. Manually harvesting this "dark data" is a labor-intensive, error-prone, and costly endeavor. This research presents an automated pipeline designed to efficiently extract, standardize, and validate Sm-Nd isotopic data from the geoscientific literature, transforming unstructured PDF tables into a machine-readable global database.
Methodology
The developed workflow integrates computer vision and domain-specific geochemical logic across several stages:
Document Retrieval and Filtering
The process begins with metadata extraction using the CERMINE tool. Over 20,000 publications were screened using targeted keywords such as "εNd" and "143Nd/144Nd" to identify relevant documents containing isotopic data.
Automated Table Extraction
The core of the pipeline utilizes a multi-step extraction process:
- Table Detection: A Fast R-CNN model, pre-trained on the TableBank dataset, identifies table coordinates with 97% accuracy.
- Structure Recognition: To accommodate diverse table styles, image processing techniques like morphological dilation and erosion are employed to reconstruct the table's skeleton, even in the absence of visible grid lines.
- Optical Character Recognition (OCR): For older, scanned documents lacking a text layer, EasyOCR is used to digitize numerical values.
Geochemical Standardization and Validation
Unlike general-purpose extraction tools, this pipeline includes a "geochemical engine" that recalculates isotopic values (e.g., εNd, TDM1, TDM2) using unified chondritic standards. This ensures scientific comparability across decades of literature. A "human-in-the-loop" approach was maintained, with 2,118 data points manually validated to benchmark the system's performance.
Results and Dataset Expansion
The automated pipeline successfully compiled 10,624 Sm-Nd data entries from 9,138 tables. This effort resulted in a 20% increase in the existing global sample count, adding 2,118 high-quality, verified data points. The automated method proved to be 27 times faster than manual curation, significantly reducing the bottleneck for global geological modeling. Regional subsets covering major orogenic belts, such as the Himalayas and the North American Cordillera, were also released to the community.
Discussion
The study highlights that raw data extraction is insufficient for geoscientific utility; recalculation using standardized formulas is mandatory to account for varying laboratory standards. While the pipeline achieved an 84.04% consistency rate compared to original author intent, challenges remain. Older scanned PDFs introduce noise that complicates OCR, and the lack of universal naming conventions for table headers (e.g., "Age" vs. "t (Ma)") requires sophisticated keyword dictionaries.
Conclusion and Future Work
This automated approach demonstrates the potential to rescue decades of scientific data from PDF "graveyards." Future iterations aim to integrate Large Language Models (LLMs) to better interpret non-standard headers and extract contextual information from the surrounding text. Such advancements will further enhance the resolution of crustal evolution maps and global geochemical databases.
Summary updated