Automated Extraction of Multicomponent Alloy Data Using Large Language Models for Sustainable Design

TL;DR

This work tackles the challenge of transforming unstructured literature on multicomponent high-entropy alloys into structured, machine-readable databases to enable sustainable design. It presents a two-stage LLM-based extraction pipeline: QS1 harvests text-derived alloy information from abstracts and methods, while QS2 mines tabular data with a curated 354-property vocabulary, guided by retrieval-augmented generation and prompting strategies. The approach yields two large databases (DB1: 37,711 text-derived records; DB2: 148,069 table-derived records) from 10,829 articles, with strong evaluation metrics and extensive post-processing to ensure composition standardization. The mined data support sustainability-informed screening across lightweighting, soft magnetics, and corrosion resistance, identifying several candidate alloys that balance performance with lower environmental and supply risks; datasets are publicly available to enable broader reuse and further development in sustainable materials design.

Abstract

The design of sustainable materials requires access to materials performance and sustainability data from literature corpus in an organized, structured and automated manner. Natural language processing approaches, particularly large language models (LLMs), have been explored for materials data extraction from the literature, yet often suffer from limited accuracy or narrow scope. In this work, an LLM-based pipeline is developed to accurately extract alloy-related information from both textual descriptions and tabular data across the literature on high-entropy (or multicomponent) alloys (HEA). Specifically two databases with 37,711 and 148,069 entries respectively are retrieved; one from the literature text, consisting of alloy composition, processing conditions, characterization methods, and reported properties, and other from the literature tables, consisting of property names, values, and units. The pipeline enhances materials-domain sensitivity through prompt engineering and retrieval-augmented generation and achieves F1-scores of 0.83 for textual extraction and 0.88 for tabular extraction, surpassing or matching existing approaches. Application of the pipeline to over 10,000 articles yields the largest publicly available multicomponent alloy database and reveals compositional and processing-property trends. The database is further employed for sustainability-aware materials selection in three application domains, i.e., lightweighting, soft magnetic, and corrosion-resistant, identifying multicomponent alloy candidates with more sustainable production while maintaining or exceeding benchmark performance. The pipeline developed can be easily generalized to other class of materials, and assist in development of comprehensive, accurate and usable databases for sustainable materials design.

Figures (6)

• Figure 1: Overall LLM-based alloy data extraction pipeline developed in this work. Starting from list of article DOIs, different sections of the article were separately extracted in XML format using publisher APIs. This data was selectively passed through two LLM extractor consisting of two query sets. While query set 1 (QS1) processed textual data from abstract and experimental section of the article, query set 2 (QS2) processed tabular data. Each of the query sets have their own series of prompts involving context, few-shot examples, etc. to finally output information such as alloy composition, properties, among others, in a tabular format. Mined data entries from each article were appended to associated databases, which were further cleaned to develop the alloy datasets. Finally, the QS2 database containing diverse property records of alloy systems was used for sustainable materials selection across 3 different application domains.
• Figure 2: a) Schematic of the QS1 prompt structure, comprising system and formatting instructions, domain-specific context, RAG-selected few-shot examples, and user query with the target article paragraph, organized in a chain-of-thought framework. b) QS1 data extraction workflow for a single article, in which alloy compositions are first extracted from each input paragraph, followed by extraction of the corresponding processing, characterization and property data, all of which are subsequently appended into DB1.
• Figure 3: a) Extended confusion matrix used in this work to evaluate QS1 and QS2. An example review article and corresponding LLM extracted datasets, capturing different types of relevant ($\mathcal{R}$) and retrieved ($\mathcal{A}$) entries. Venn diagrams capturing evaluation of b) QS1 on review article dataset, and of QS2 on c) review article and d) manually curated datasets.
• Figure 4: Overall framework for QS2. a) First, a narrowed list of property set were identified in the target article table using a modified RAG approach and a master list of 354 materials properties constructed using DB1. b) This selected property list was included as context in the QS2 prompt, which also included other aspects of prompt engineering such as chain-of-thought prompting, RAG-based few-shot examples and formatting instructions. An example few-shot example for QS2 is also included. c) Data records output by QS2 were post-processed for any formatting errors before being appended to DB2. See main text for more details on QS2 workflow.
• Figure 5: Visualization of the LLM-mined DB2. a) Periodic-table heatmap showing the frequency of occurrence of elements in the database. b) Log-scale distribution of the 20 most frequently extracted material properties. t-SNE visualization of the full database, with colors denoting c) hardness, d) yield strength, e) density and f) family of the alloy.