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

This paper presents an LLM-based pipeline that accurately extracts multicomponent alloy data from both text and tables to create the largest publicly available database of its kind, enabling sustainable materials design by identifying high-performance alloy candidates for lightweighting, soft magnetic, and corrosion-resistant applications.

The Gist

Imagine the world of materials science as a massive, chaotic library containing millions of books. These books describe how to make new, super-strong, or eco-friendly metal alloys (mixtures of metals). The problem is that the information inside is messy. Some facts are hidden in paragraphs of text, others are buried in complex tables, and the way scientists write about them varies wildly. One scientist might call a metal "Al-HEA," while another writes out a long chemical formula. Trying to find the best recipe for a specific job by reading these books one by one is like trying to find a single specific grain of sand on a beach by hand—it's slow, tedious, and impossible to do at scale.

This paper introduces a solution: a team of super-smart AI robots (called Large Language Models, or LLMs) that act as automated librarians. Their job is to read these thousands of scientific papers, understand the messy information, and organize it into a clean, searchable digital database.

Here is how they did it, broken down into simple steps:

1. The Two-Step Cleaning Process

The researchers realized they couldn't just ask the AI to "read everything." They needed a strategy, so they built a two-stage pipeline:

• Stage 1: The "Skimmer" (Text Extraction)
  First, the AI reads the abstracts and the "how we made it" sections of the papers. Think of this as skimming the back of a cereal box to see what ingredients are listed. The AI looks for:

  • What metals are in the mix?
  • How was it heated or cooled?
  • What tests were run on it?
  • Result: They built a database with 37,711 entries just listing the recipes and the types of tests used.

• Stage 2: The "Deep Diver" (Table Extraction)
  Next, the AI dives into the tables where the actual numbers live. This is harder because tables are tricky. A column might say "Hardness" in one paper and "HV" in another. The AI had to be taught to recognize that these mean the same thing. It extracted the specific numbers (like "500 MPa") and the conditions (like "at 20 degrees Celsius").

  • Result: They built a second, even larger database with 148,069 entries containing the actual performance numbers.

2. Teaching the AI to Be an Expert

You can't just ask a generic AI to read science papers; it might get confused or make things up (a problem called "hallucination"). To fix this, the researchers used a technique called Prompt Engineering.

Think of this as giving the AI a specialized instruction manual before it starts working. They told the AI:

• "You are a materials science expert."
• "Here is a dictionary of how metals are named."
• "Here are 98 examples of how to read a sentence and pull out the right numbers."
• "If you aren't sure, say 'I don't know' instead of guessing."

They also used a trick called RAG (Retrieval-Augmented Generation). Imagine the AI is taking a test. Instead of relying only on its memory, it has a cheat sheet. Before answering a question about a specific alloy, the AI looks up similar examples from its training data to see how an expert would answer that specific type of question. This made the AI much more accurate.

3. The Result: A Giant, Clean Database

By applying this system to over 10,000 scientific articles, the team created the largest publicly available database of multicomponent alloys (often called High-Entropy Alloys).

• They found that the AI was about 83% to 88% accurate, which is as good as or better than previous methods.
• They cleaned up the data so that "Al-HEA" and "Aluminum High Entropy Alloy" are now understood as the same thing.

4. Putting the Database to Work: The "Green" Test

The researchers didn't just stop at building the library; they used it to solve a real-world problem: Sustainability.

They wanted to find alloys that are not only strong but also good for the planet. They looked at three specific jobs:

1. Lightweighting: Making cars and planes lighter to save fuel.
2. Soft Magnetism: Making better motors and transformers for electricity.
3. Corrosion Resistance: Making materials that don't rust in saltwater or chemicals.

They combined the performance data (how strong is it?) with a "Sustainability Score" (how hard is it to mine these metals? How much pollution does making them cause?).

The Discovery:
They found several new alloy recipes that are better than the current commercial metals used today. These new alloys are not only strong or rust-resistant but are also made from elements that are more abundant and easier to recycle, making them a greener choice for the future.

Summary

In short, this paper is about using AI as a super-powered translator and organizer. It took a mountain of messy, unstructured scientific writing and turned it into a clean, organized spreadsheet. This new spreadsheet allows scientists to quickly find the best, most eco-friendly metal recipes for specific jobs, speeding up the invention of sustainable materials. The team has made this database and the code they used available to everyone online so others can use it too.

Technical Summary

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

Problem Statement
The design of sustainable materials requires the simultaneous optimization of performance and environmental impact. However, the vast knowledge generated by decades of experimental and computational studies on multicomponent alloys (specifically high-entropy alloys or HEAs) remains dispersed across scientific literature in unstructured formats (text, tables, figures). Manual organization of this data is infeasible at scale. While previous automated extraction efforts utilizing task-specific machine learning (ML) models or rule-driven methods have shown promise, they often struggle with the complex, context-dependent relationships inherent in materials science (e.g., linking a property value to specific testing conditions). Furthermore, early Large Language Model (LLM) applications for data extraction suffered from limited accuracy, narrow property scopes, and a lack of standardized compositional data, hindering their utility for constructing comprehensive, machine-readable databases required for sustainable materials selection.

Methodology
The authors developed a two-stage, LLM-based pipeline to extract structured data from over 10,000 research articles on multicomponent alloys. The workflow leverages OpenAI's GPT-4o and GPT-4o mini models, enhanced by advanced prompt engineering techniques including system-level instructions, domain-specific context, few-shot learning, chain-of-thought reasoning, and Retrieval-Augmented Generation (RAG).

1. Literature Corpus Construction: Approximately 18,000 articles were queried from Web of Science. 10,829 accessible articles were retrieved via the Elsevier API, parsed from XML, and segmented into abstracts, experimental sections, and tables.
2. Query Set 1 (QS1) – Textual Extraction:
   • Target: Extract alloy systems, processing conditions, characterization techniques, and reported property names from abstracts and experimental sections. Numerical values were not extracted at this stage.
   • Technique: A RAG-based approach retrieved the top-5 most similar expert-annotated examples from a vector database (containing 98 curated examples across four categories) to guide the LLM. The prompt utilized a chain-of-thought structure to guide the model through task definition, context understanding, and extraction.
   • Output: A database (DB1) of alloy-level records defined by composition and processing conditions.
3. Query Set 2 (QS2) – Tabular Extraction:
   • Target: Extract quantitative property values, units, testing conditions, and alloy compositions from tables.
   • Technique: A two-step LLM process was employed. First, a modified RAG approach identified property headers in tables by matching them against a master list of 354 standardized material properties (derived from DB1). Second, a targeted prompt extracted the numerical values and associated conditions for the identified properties.
   • Post-Processing: A multi-stage pipeline involving rule-based checks and a secondary LLM call was used to standardize non-standard nomenclature (e.g., converting "SS304" to chemical compositions), increasing the number of usable entries.
4. Sustainability Integration: The resulting databases were coupled with nine curated sustainability indicators (covering supply risk, environmental burden, and socio-economic factors) to evaluate alloy candidates across three application domains.

Key Contributions

• Dual-Database Construction: The pipeline generated two complementary databases:
  • DB1: 37,711 entries containing alloy compositions, processing conditions, and property names.
  • DB2: 148,069 entries containing quantitative property values, units, and experimental conditions.
• Standardized Property Vocabulary: The authors curated a controlled vocabulary of 354 distinct material properties to ensure consistency, while including an "Others" category to capture non-standard or article-specific metrics without discarding data.
• RAG-Enhanced Prompting: The integration of RAG to dynamically select few-shot examples based on semantic similarity significantly improved extraction accuracy and domain sensitivity compared to static prompting.
• Public Accessibility: The curated datasets are made publicly available via an interactive web platform, "Alloy Tattvasar," along with source code and prompt templates on GitHub.

Results

• Extraction Performance:
  • QS1 (Text): Achieved an F1-score of 0.83 against expert-annotated benchmarks from review articles. The system demonstrated high precision in identifying alloy compositions but faced challenges with complex multi-step arithmetic substitutions in chemical formulas.
  • QS2 (Tables): Achieved an F1-score of 0.88 across a broad range of properties (mechanical, corrosion, electrical, etc.), with a notably high F1-score of 0.96 for mechanical properties. Precision was near-perfect (0.98), though recall was limited (~80%) due to the complexity of tabular data and strict instructions to omit uncertain entries.
• Database Statistics: The final DB2 contained 16,381 unique alloy compositions. Visualization via t-SNE revealed distinct clustering of alloy families and correlations between properties (e.g., hardness and yield strength).
• Sustainable Design Case Studies: The database was applied to three domains:
  1. Lightweighting: Identified 9 HEA candidates (e.g., Fe-rich, Al/Ti-rich systems) with competitive specific strength and favorable sustainability indices compared to commercial aluminum and titanium benchmarks.
  2. Soft Magnetism: Identified 11 candidates (e.g., Al-Co-Cr-Fe-Ni-Si systems) offering performance comparable to FINEMET and Permalloy but with improved sustainability by reducing reliance on critical elements like Ni and Co.
  3. Corrosion Resistance: Identified candidates (e.g., Cr-Fe-Al-Ni systems) with corrosion current densities comparable to SS 316L but with higher sustainability indices due to the absence of Molybdenum.

Significance and Claims
The paper claims that this work demonstrates the capability of LLM-based methods to transform complex, unstructured materials literature into structured, accurate, and usable databases suitable for downstream AI/ML and materials discovery tasks. The authors emphasize that their approach successfully captures not just property values but the critical experimental context (processing and testing conditions) required for meaningful data interpretation.

The study highlights that the developed pipeline is generalizable to other material classes and can be coupled with rigorous sustainability methodologies (such as Life Cycle Assessment) to support sustainable materials design. The authors acknowledge current limitations, specifically the exclusion of data embedded in figures and the challenges in fully standardizing microstructural information for multiphase systems, noting that future advancements in LLMs and agentic frameworks are expected to address these gaps. The work positions the resulting database as a foundational resource for learning composition-property relationships and accelerating the discovery of sustainable multicomponent alloys.