Construction of a Battery Research Knowledge Graph using a Global Open Catalog

This paper presents a scalable pipeline for constructing an author-centric knowledge graph of battery research using OpenAlex and LLM-enhanced keyphrase extraction to enable cross-institutional expertise tracking, community detection, and semantic similarity analysis beyond traditional citation-based methods.

The Gist

Imagine the world of battery research as a massive, chaotic library containing nearly 190,000 books (scientific papers). In this library, thousands of authors are writing about everything from the chemistry of lithium to the safety of manufacturing. The problem? It's incredibly hard to find the right person to collaborate with because the books are scattered, the titles are confusing, and the authors are hidden behind institutional walls.

This paper describes a project to build a smart, interactive map of this library, but instead of organizing it by books, it organizes it by people.

Here is how they did it, explained through simple analogies:

1. The Starting Point: The "Big Book" (OpenAlex)

The researchers started with a giant, free digital catalog called OpenAlex. Think of this as a massive library card catalog that already has some basic tags on the books (like "Physics" or "Energy").

• The Problem: These tags are too broad. If you tag a book just "Science," it doesn't tell you if the author is an expert in batteries or astronomy.
• The Fix: They needed to read the actual pages (titles and abstracts) to find the specific, detailed topics.

2. The Detective Work: Finding the "Secret Keywords"

To understand what each author really cares about, the team used a digital detective named ChatGPT (specifically a version called gpt-3.5-turbo).

• The Analogy: Imagine you have a stack of 190,000 resumes. You ask ChatGPT to read the "About Me" section of each one and pull out the top 2-3 most important skills.
• The Result: Instead of just seeing "Science," the system now sees specific phrases like "solid-state electrolytes" or "lithium-ion degradation." They tested different "detectives" (AI models) and found that ChatGPT was the best at understanding the nuance of battery language.

3. Creating the "Author ID Card" (The Vector)

Once they had the keywords, they created a unique ID card for every single researcher. This isn't a photo ID; it's a weighted scorecard.

• How it works: Imagine a recipe for a smoothie.
  • Freshness: If an author published a paper yesterday, it counts for more than a paper from 1995 (because science moves fast).
  • Leadership: If the author was the first person to write the paper (the main chef), their ideas count more than if they were just a helper on the team.
  • Ingredients: They mix the broad tags from the library catalog with the specific keywords found by the AI.
• The Output: Every author gets a mathematical "fingerprint" that represents their unique research style and current focus.

4. The Magic Map: Finding Your Tribe

With these ID cards, the researchers built a browser-based map (a Knowledge Graph).

• The Visual: Think of a galaxy where every star is a scientist.
  • Blue Lines: Connect scientists who are very similar (they use the same keywords and ideas). These are your potential best friends or collaborators.
  • Yellow Lines: Connect scientists who are indirectly related (maybe you know someone who knows them). These are potential "weak ties" that could lead to new opportunities.
• The Search: You can type in a name, and the map zooms in to show you who they know and what they are working on. You can also search by a topic (like "anodes") and see a cloud of words showing who is the expert in that area.

5. The "Universal Translator" (RDF & Wikidata)

Finally, they didn't just keep this map for themselves. They translated it into a universal language called RDF and linked it to Wikidata (like Wikipedia's database).

• The Analogy: Imagine they took their custom map and gave it a universal passport. Now, other computer systems, libraries, or even different scientific fields (like solar energy or AI) can "speak" to this battery map and understand it. It breaks down the walls between different databases.

Why Does This Matter?

Before this, finding a collaborator was like trying to find a needle in a haystack by guessing.

• Old Way: "I know a guy at University X who works on batteries."
• New Way: "I need an expert in 'solid-state safety' who is currently active. Here is a map showing the top 10 people in the world who match that exact description, regardless of which university they work for."

In short: They built a smart, AI-powered GPS for battery scientists, helping them find their perfect research partners across the globe, based on what they actually write and do, not just where they work.

Technical Summary

1. Problem Statement

Battery research is a rapidly expanding, highly interdisciplinary field intersecting materials science, electrochemistry, and manufacturing. The sheer volume of publications and the fragmentation of knowledge across institutional boundaries make it difficult to:

• Track relevant expertise and terminology.
• Identify potential collaborators across institutions.
• Perform author-level discovery based on domain semantics rather than just citation networks or co-authorship structures.

Existing solutions are often organized around documents or entities (ontologies) rather than authors, limiting their utility for quantitative similarity retrieval and expertise identification at a cross-institutional scale.

2. Methodology

The authors propose a pipeline to construct an author-centric knowledge graph using the OpenAlex open bibliographic catalogue. The workflow consists of four main stages:

A. Dataset Preparation

• Source: Data was retrieved from OpenAlex using the identifier for "Battery (electricity)" (C555008776).
• Scale: The initial corpus contained 189,581 works corresponding to 356,103 authors.
• Filtering: Records older than 1990 or lacking publication dates were removed. For the proof-of-concept demonstration, the dataset was refined to the top 10,000 authors by publication count to reduce computational load.
• Preprocessing: Generic "battery" concepts were pruned, and concepts with zero confidence scores were removed to reduce noise.

B. Hybrid Research Descriptor Extraction

To overcome the coarse granularity of OpenAlex's pre-assigned concepts, the authors employed a hybrid approach:

1. OpenAlex Concepts: Used as coarse-grained domain tags.
2. Keyphrase Extraction: Fine-grained keyphrases were extracted from titles and abstracts using KeyBERT.
   • Model Selection: Four models were benchmarked (SentenceTransformers, domain-tuned BERT variants, and ChatGPT). ChatGPT (gpt-3.5-turbo) was selected as the backend model, achieving the highest similarity score (0.6781) against a silver-standard ground truth derived from Grobid.
   • Extraction Limits: Up to 2 keyphrases from titles and 10 from abstracts were retained per paper.

C. Author Vector Construction

A weighted research descriptor vector ($\vec{v}_a$) was generated for each author, aggregating their entire publication history. The vector construction incorporated three specific weighting factors:

1. Descriptor Origin: Different weights were applied to OpenAlex concepts versus extracted keyphrases.
2. Temporal Recency: Publications were divided into three periods (1990–2000, 2001–2010, 2011–2023). A decay factor ($f_{pj}$) was applied to older periods to emphasize current expertise.
3. Authorship Position: First-author publications were weighted more heavily ($w_{first}$) than non-first-author publications ($w_{nonfirst}$), as they better reflect the primary research focus.

The final vector is calculated as:
$$\vec{v}_a = \sum_{j=0}^{2} f_{pj} \cdot (w_{first} \cdot \vec{D}_{pj} + w_{nonfirst} \cdot \vec{D}'_{pj})$$
Where $\vec{D}$ represents the vector of descriptors (concepts and keyphrases) weighted by frequency and relevance scores.

D. Graph Construction and Visualization

• Similarity Computation: Author–author similarity was computed based on the cosine similarity of their research descriptor vectors.
• Visualization:
  • Word Clouds: Generated to represent individual author profiles, with font size proportional to descriptor relevance.
  • Interactive Map: A browser-based interface allows users to search by author or concept. It visualizes direct connections (shared descriptors) and indirect connections (co-authorship chains).
• Serialization: The graph was serialized into RDF (Resource Description Framework) and linked to Wikidata identifiers to ensure interoperability with external Linked Open Data sources.

3. Key Contributions

• Author-Centric Framework: Unlike prior works focused on document clustering or institutional repositories, this approach creates a cross-institutional, author-level representation grounded in domain semantics.
• Hybrid Semantic Representation: The novel combination of coarse OpenAlex concepts with fine-grained LLM-extracted keyphrases provides a richer, more nuanced view of researcher expertise than either method alone.
• Weighted Vector Model: The introduction of a multi-factor weighting system (temporal decay, authorship position, and descriptor origin) allows for dynamic and explainable similarity matching.
• Open Interoperability: By serializing the graph in RDF and linking to Wikidata, the system is designed to be extensible beyond the battery domain and integrable with broader scientific knowledge graphs.

4. Results

• Dataset: Successfully processed a corpus of ~190k battery-related papers.
• Model Performance: The evaluation of keyphrase extraction showed that ChatGPT (gpt-3.5-turbo) outperformed specialized BERT variants and SentenceTransformers in semantic similarity to ground-truth keywords.
• Visualization: The system successfully generated interactive similarity maps and word clouds.
  • Example: The map visualized a researcher's direct connections (blue arcs) based on shared descriptors and indirect connections (yellow arcs) based on co-authorship, effectively highlighting potential new collaborators.
• Interoperability: The RDF serialization successfully linked author nodes to Wikidata, enabling standardized querying and integration with external datasets.

5. Significance

This work addresses a critical gap in materials informatics by providing a scalable, semantic tool for navigating the battery research landscape.

• Collaboration Discovery: It enables researchers to find collaborators based on specific technical expertise (e.g., "lithium-ion anodes") rather than just institutional proximity or citation counts.
• Domain Adaptability: While applied to batteries, the pipeline is domain-agnostic and can be adapted to other scientific fields.
• Future-Proofing: The use of open catalogs (OpenAlex) and standard formats (RDF) ensures the knowledge graph remains updatable and interoperable.
• Limitations & Future Work: The authors note that future iterations could utilize open-weight LLMs for better transparency, implement stricter merging of semantically overlapping keyphrases (e.g., "anode" vs. "anode material"), and adjust weighting factors to better represent early-career researchers.