Harnessing X-ray Absorption Spectroscopy Data through Multimodal Mining of Battery Literature

X-ray absorption spectroscopy (XAS) is a critical tool for probing the local electronic and atomic structure of materials. Despite its importance, the vast majority of published XAS data remains trapped in static image files within scientific papers. This unstructured format prevents researchers from performing large-scale data-driven analyses, such as training machine learning models for autonomous materials discovery or conducting cross-laboratory comparisons.

The authors created an automated, multimodal pipeline to convert these literature-embedded figures into structured, machine-readable data. The process begins by acquiring full-text articles and using a multi-stage filtering approach—combining large language models (LLMs) and vision language models (VLMs)—to identify relevant XAS figures.

Once identified, the pipeline employs a specialized curve-segmentation algorithm that uses color decomposition and connectivity graphs to isolate individual spectral traces, even when they overlap. To ensure the data is interpretable, the system uses OCR and VLMs to extract axis labels and tick marks, allowing for the precise conversion of pixel coordinates into physical energy and absorption units. Finally, the pipeline links each spectrum to its corresponding metadata (such as the absorbing element, absorption edge, and material composition) by synthesizing contextual information from figure captions and the surrounding text.

Applying this pipeline to over 485,000 battery-related papers, the researchers successfully extracted 13,740 high-quality XAS curves. This dataset covers 66 different absorbing elements and a wide range of battery chemistries. By providing this data in a structured JSONL format, the authors lower the barrier for high-throughput characterization and provide a foundation for future AI-driven materials science research.

X-ray Absorption Spectroscopy (XAS)

An experimental technique used to determine the local atomic and electronic structure of materials by measuring the absorption of X-rays as a function of energy.

Multimodal Mining

A data extraction approach that simultaneously analyzes different types of information, such as images (figures) and text (captions and body paragraphs), to reconstruct structured data.

Vision Language Model (VLM)

An AI model capable of interpreting and reasoning about visual information, used here to classify figures and extract textual information from plots.

Curve Segmentation

The computational process of isolating individual data lines from a plot, particularly when multiple lines overlap or intersect.

Edge (Absorption Edge)

A specific energy level in an X-ray absorption spectrum corresponding to the ionization of a core-level electron of a specific element.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.