matscholar

matscholar extracts structured materials science information from scientific publications using natural language processing and named entity recognition to enable programmatic querying and literature-scale analysis.

Key Features:

• Implementation: Python library for natural language processing and text mining in the materials science domain.
• Information extraction: Named entity recognition (NER) models that convert unstructured text into structured database entries.
• Entity types: Extraction of inorganic material mentions, sample descriptors, phase labels, material properties and applications, and synthesis and characterization methods.
• Scale: Applied to a corpus of 3.27 million materials science abstracts, yielding over 80 million extracted named entities.
• Performance: NER classifier reported with an F1 score of 87%.

Scientific Applications:

• Literature-scale meta-analysis: Enable aggregated analysis across millions of abstracts to characterize trends and patterns in materials research.
• Programmatic querying: Produce structured database entries to support automated queries over materials mentions, properties, and methods.
• Knowledge synthesis for materials discovery: Facilitate extraction of synthesis, characterization, and property information to inform hypothesis generation and comparative studies.

Methodology:

Uses Python-based NLP and text-mining with a trained named entity recognition model to identify specified entity types (inorganic material mentions, sample descriptors, phase labels, material properties and applications, synthesis and characterization methods), transforming abstracts into structured database entries; applied at scale to 3.27 million abstracts with an NER F1 score of 87% and over 80 million extracted entities.