Large language model-enabled automated data extraction for concrete materials informatics

This paper introduces a generalizable, LLM-powered pipeline that automates the extraction of high-quality experimental data from unstructured scientific literature, successfully constructing the largest open database for blended cement concrete to accelerate materials informatics.

The Gist

The Problem: The "Library of Babel" for Concrete

Imagine you are a world-class chef trying to invent the perfect, most eco-friendly recipe for a new kind of bread. You know that thousands of secret recipes exist, but there’s a massive problem: they aren't in a neat cookbook. Instead, they are scattered across millions of handwritten notes, messy napkins, and old, dusty journals hidden in different libraries all over the world.

Some notes use grams, some use ounces; some call it "flour," others call it "wheat powder." To find the best recipe, you’d have to hire a thousand assistants to read every single note, translate the units, and type them into a spreadsheet. It would take lifetimes.

This is exactly the problem scientists face with concrete. Concrete is the "bread" of our modern world—it’s everywhere, but making it is a huge source of CO2 pollution. To make it greener, we need to study thousands of old experiments to see which "ingredients" (like recycled ash or clay) work best. But all that data is trapped inside millions of scientific papers, buried in messy tables and confusing text.

---

The Solution: The "Super-Librarian" (The LLM Pipeline)

The researchers at Rice University decided to stop hiring human assistants and instead built a "Super-Librarian."

They used Large Language Models (LLMs)—the same kind of "brains" that power ChatGPT—and organized them into a specialized team of digital experts. Instead of one giant robot trying to do everything, they created a relay race of specialized agents:

1. The Scout: This agent flies through the digital library, scanning millions of papers to find only the ones that actually talk about concrete recipes.
2. The Table Specialist: This agent looks at messy, complicated tables (the kind with merged cells and weird headers) and "reads" them, turning them into clean, digital lists.
3. The Translator: This agent is the master of units. If one paper says "pounds" and another says "kilograms," or if one uses a weird nickname for an ingredient, the Translator fixes it so everything matches perfectly.
4. The Fact-Checker: Finally, a specialized agent checks the math. If a recipe says the concrete has "negative weight" (which is impossible!), the Fact-Checker flags it as an error.

---

The Result: A Massive Digital Cookbook

This "Super-Librarian" team worked incredibly fast. In just one hour, they scanned through 27,000 scientific papers and extracted nearly 9,000 high-quality "recipes" (data records).

Before this, the best "cookbooks" scientists had were tiny—some only had a few hundred recipes. This new database is the largest, most detailed "digital cookbook" for concrete ever created. It doesn't just list the ingredients; it lists the temperature of the room, how long the concrete sat there, and exactly how strong it became.

---

Why Does This Matter? (The "Smart Oven")

Now that we have this massive, clean database, we can feed it into Machine Learning—think of this as a "Smart Oven."

Because the "Smart Oven" has seen 9,000 different recipes, it can start to predict the future. A scientist can say, "Hey, if I mix 20% recycled clay with 10% fly ash and bake it at 70 degrees, how strong will it be?"

Instead of spending months in a lab doing trial-and-error (which wastes time and money), the computer can give a highly accurate answer in seconds. This allows us to design new, ultra-strong, and low-carbon concrete much faster, helping us fight climate change by building a greener world, one "recipe" at a time.

Technical Summary

Technical Summary: Large Language Model-Enabled Automated Data Extraction for Concrete Materials Informatics

1. Problem Statement

The advancement of data-driven materials discovery is severely hindered by the "data scarcity bottleneck." While computational datasets are abundant, experimental datasets—which are essential for real-world validation—are often small, proprietary, or fragmented.

In the field of concrete materials informatics, this problem is particularly acute due to:

• Compositional Complexity: Unlike single-element or simple compound materials, concrete is a composite mixture of multiple constituents (binders, aggregates, admixtures), each with distinct chemical and physical properties.
• Heterogeneous Reporting: Scientific literature presents data in highly unstructured formats, including complex tables with multi-level headers, inconsistent units, non-standard naming conventions, and domain-specific acronyms.
• Information Fragmentation: Critical attributes (e.g., curing conditions, specimen dimensions, and chemical compositions) are often scattered across different sections of a paper (text vs. tables) or even across multiple tables.

2. Methodology

The authors propose a generalizable, LLM-powered automated data extraction pipeline designed to transform unstructured scientific literature into structured, machine-readable databases. Using blended cement systems as a representative case, the pipeline employs a sequential chain of specialized LLM-based agents:

• Article Retrieval & Screening: A large corpus was screened using an LLM-assisted approach to identify relevant publications (focusing on blended cement systems) from a pool of over 27,000 articles.
• Extraction Agents:
  • Table Extraction Agents: Parse complex XML/HTML tables (handling merged cells and multi-level headers) into a structured plain-text format.
  • Text Extraction Agents: Use regular expressions and LLMs to identify and extract information from prose, such as acronym definitions, curing conditions (temperature/humidity), and specimen geometry.
• Processing Agents: These agents perform critical "cleaning" and "reconciliation" tasks:
  • Normalization: Standardizing units (e.g., psi to MPa) and naming conventions.
  • Acronym Expansion: Using extracted definitions to resolve table abbreviations.
  • Data Integration: Merging information from disparate tables and text segments into a single, internally consistent record for each mixture.
  • Derived Attribute Computation: Calculating values like water-to-binder ratios or specimen volumes.
• Validation: The pipeline was benchmarked against a human-extracted "ground-truth" dataset of 3,015 records from 58 publications.

3. Key Contributions

• A Modular LLM Pipeline: A scalable framework that decomposes the complex task of scientific data mining into manageable sub-tasks handled by specialized agents.
• The Largest Open Concrete Database: The construction of a high-quality, open-access database containing 8,979 unique records with over 100 attributes (including chemical oxide compositions and physical properties).
• Benchmarking of LLMs: A comprehensive evaluation of 17 different LLMs (both open and proprietary), demonstrating that the pipeline's success is not dependent on a single model but is a result of the architectural design.

4. Results

• High Accuracy: The pipeline achieved exceptional performance, with an overall F1 score of up to 0.97 (using Claude Sonnet 4.5). Even the widely used GPT-4o achieved an overall F1 score of 0.95.
• Efficiency & Scalability: The pipeline processed the entire curated corpus of 278 publications in less than one hour using parallelization, representing a massive speedup compared to human experts (who would require ~25 hours for the same task).
• Machine Learning Insights:
  • In-Distribution (ID) Accuracy: Including chemical/physical descriptors (like oxide composition) improved the predictive accuracy of compressive strength models.
  • Out-of-Distribution (OOD) Generalization: The study proved that larger, more diverse datasets and the inclusion of detailed binder descriptors significantly improve a model's ability to predict properties for "unseen" materials (e.g., moving from binary to quaternary blends).
  • Scaling Laws: Predictive error (RMSE) followed a power-law relationship with training data size, confirming the necessity of large-scale data for high-performance ML.

5. Significance

This work represents a paradigm shift in how materials informatics datasets are constructed. By moving away from manual curation toward automated, LLM-driven pipelines, the research provides a scalable solution to the data scarcity problem.

The implications are twofold:

1. For Concrete Science: It provides the foundational data necessary to design more sustainable, low-carbon concrete (e.g., optimizing SCM replacements to reduce CO2 emissions).
2. For Materials Informatics at Large: The modular design is domain-agnostic. It can be readily adapted to any materials field (metals, polymers, ceramics) where data is trapped in the "unstructured" medium of scientific literature, thereby accelerating the global transition toward data-driven materials discovery.